DEV Community: Gajanand V Dhayagode

My AI field guide called castor a "Virginia Creeper". A 107-line list fixed more than fine-tuning did.

Gajanand V Dhayagode — Sun, 11 Oct 2026 17:48:05 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 1 — Touch Grass.

What I Built

Pocket Naturalist is an offline field guide for walks. You photograph a plant, bird or bug, and Gemma 4 12B, running on my own laptop, tells you what it probably is. It also points out one detail to look at more closely, gives you a small quest to find next ("find a plant with compound leaves"), and writes a journal of your walk at the end.

There's no cloud and no account, and it works with no mobile signal: the phone talks to the laptop over a hotspot. You can snap photos live, or, if you don't want to carry a laptop around a garden (I didn't), take photos with your normal camera and add them from the gallery afterwards.

The rule I gave myself: the screen should be the shortest part of the walk. The answer is read aloud, and the phone goes back in your pocket.

What it does:

👀 Shows its reasoning: the features it saw, its best guess, and what else it could be.
🔁 Takes a second look when it's unsure: it describes the plant's structure first, then names it.
🔍 Mystery finds: when it's still unsure, it gives you one thing to check on the real plant ("blue flowers mean butterfly pea…") instead of a confident wrong answer.
✏️ "I know what this is": correct it, and the journal gives you the credit.
🌿 An Indian field guide: the fix that made it actually useful where I live. More on that below.

Demo

Two weeds Gemma got wrong before the field guide. Notice the "look closer" tips: they only describe what's actually in the photo.

It was "Sago Palm" before the field guide.

When Gemma isn't sure, it doesn't guess. It tells you what to check on the real plant.

The walk journal, written by Gemma from my 17 finds. It uses Gemma's own names, mistakes included (that's not teak), which is exactly why the app shows "could also be" and lets you correct it.

On my real 17-photo walk, here's what changed after the fix:

Plant	Stock Gemma 4	With the field guide
Castor	Virginia Creeper	✅ Castor
Curry leaf	Chinese Fringe Tree	✅ Curry leaf
Congress grass	Common Ragweed	✅ Parthenium
Four o'clock flower	"Mexicaner" (not a plant)	✅ Four o'clock flower
Black nightshade	Wild Jasmine	✅ Black nightshade
Araucaria	Sago Palm	✅ Norfolk Island / Cook pine
Red rose	Adenium	✅ Rose

Overall: 4 → 10 correct out of 15 photos where I know the right answer.

Code

gajanand27-05 / pocket-naturalist

Offline field guide for walks: snap a plant or bird, local Gemma 4 identifies it and sends you on the next quest. Hacktoberfest 2026 Touch Grass.

🌿 Pocket Naturalist

An offline field guide for walks. Snap a plant, bird or bug on your phone. Gemma 4, running on a laptop in your backpack, tells you what it probably is, what to look at more closely, and gives you a small quest to find next. When you finish, it writes a journal of your walk.

You don't need mobile signal, a cloud API or an account. Your photos never leave your own devices.

Built for the Hacktoberfest 2026 Open-Source AI Challenge, Week 1: "Touch Grass".

How it works

phone (browser) ──same Wi-Fi or hotspot, no internet──▶ laptop: server.py ──▶ Ollama ──▶ gemma4:12b (vision)

Start a walk. You get your first quest, read aloud by the phone's built-in offline voice.
📷 Snap a find, or 🖼️ add photos from your gallery later. You don't have to carry the laptop: take photos in the garden with…

View on GitHub

It's one standard-library Python file plus one HTML page, with nothing to pip install:

ollama pull gemma4:12b
python server.py   # prints a URL to open on your phone

How I Built It

The setup: phone browser → Wi-Fi hotspot (no internet) → server.py on my laptop → Ollama → gemma4:12b. The laptop is nothing special: an RTX 3050 with 6 GB of GPU memory and 16 GB of RAM. Gemma 12B doesn't fully fit, so it runs partly on the GPU and partly on the CPU, at about 30–40 seconds per photo. Every answer is forced into a fixed JSON format with Ollama's format option, so the app never has to guess what the model meant.

Day 1: making it honest

The first thing I learned is that a vision model can be confidently wrong. A house sparrow came back as "Jungle Babbler, 95% sure." A "fun fact" about hibiscus was made up, and so were some of the scientific names. So I:

made Gemma list what it sees first, then rank three candidates, then choose;
removed fun facts and scientific names and replaced them with "look closer" tips that may only describe what's visible in the photo;
told the journal to use only the walk data. Its first draft described "the afternoon sun hanging low" on a walk it knew nothing about.

Then a vine on my railing came back as "Creeping Charlie". I went outside to look: it had bright blue flowers. Butterfly pea. That moment became a feature: when Gemma isn't sure, the app now gives you a clue to check on the real plant instead of a guess. That sends you back to the plant instead of back to the screen, which is the whole point of "Touch Grass".

Day 2: the real walk, and a bias I didn't expect

I walked my neighbourhood and took 17 photos of ordinary plants: castor, moringa, curry leaf, a four o'clock bush, roadside weeds. Stock Gemma got 4 of 15 right.

The interesting part was how it was wrong. For the castor plant, Gemma described "palmately lobed leaves, deeply serrated edges, large star-shaped leaf". That's a perfect description of castor. Then it named it "Virginia Creeper", an American vine. It did the same everywhere: moringa became "Purslane", curry leaf became "Chinese Fringe Tree", and a four o'clock flower became "Mexicaner", which isn't a plant at all.

Its eyes were fine. Its vocabulary was American.

Fix 1: a field guide in plain text

Real field guides are regional, so I wrote one: field_guide.csv, a list of 107 common Indian plants from gardens, kitchens, roadsides, waste ground and parks. Each entry has the features that tell it apart. For example:

Castor: shrub; very large star-shaped palmate leaves with 7–11 toothed lobes; reddish stems; spiny seed clusters

It goes into Gemma's prompt with three rules I learned the hard way:

Sort it A–Z. In my first version, "Holy basil" was the first entry, and Gemma started calling half my photos holy basil.
Plants only. Without that rule, the house sparrow became a "lark".
"Never force a match." If nothing fits, give your own best name with lower confidence.

I also had to keep the prompt short enough to fit Ollama's default 4,096-token context. When I raised it, my 16 GB laptop ran out of memory and the app server was killed. Dropping the local-language names from the prompt (they stay in the CSV for people) fixed that.

Measured on 28 labelled photos (22 of mine, 6 from the internet):

	Correct (first answer)	My garden photos
Stock Gemma 4 12B	13/28	8/22
+ field guide	20/28	16/22

I stopped tuning the list after one revision, so I wasn't just fitting it to my own test photos.

Fix 2: fine-tuning on a DGX Spark (and why it lost)

I also had access to an NVIDIA DGX Spark, so I tried the "proper" fix too: a LoRA fine-tune of Gemma 4 12B on 1,193 photos from DIMPSAR, an open dataset of Indian plant photos from Karnataka and Kerala (Pushpa B.R. & N. Shobha Rani, Data in Brief 2023, CC BY 4.0).

To keep the app's output format, each training answer was Gemma's own answer with only the species name corrected. I trained a LoRA on the language layers (rank 16, one epoch, about 70 minutes), merged it, and patched the weights into Ollama's official gemma4:12b file so it kept working with images.

(measured on the DGX)	Stock	Fine-tuned
28 test photos	12/28	13/28
DIMPSAR validation	16/100	20/100
Confidence	0.6–0.95	1.0 on every photo

+1 photo, and it became 100% sure of everything, which would switch off the app's "second look" and Mystery finds. Two lessons I'm taking with me:

Check the scale of numbers you copy into training data. 4% of Gemma's own answers said "confidence": 85 instead of 0.85. The app had quietly clamped those, so nobody noticed until the fine-tuned model said 1.0 for everything.
In structured output, the loss is mostly the long text. Validation loss fell from 1.30 to 0.33, but name accuracy only went from 16% to 20%. The species name is a few tokens next to four descriptive sentences.

So the app ships with stock Gemma 4 plus the field guide. A 107-line text file beat 70 minutes on a DGX Spark.

Other models I measured

On my first 13 test photos: Gemma 4 12B 10/13, Qwen3.5 9B 9/13, MiniCPM-V 4.5 8/13, Qwen2.5-VL 3B 8/13 (it said "100% sure" about almost everything), and Gemma 4 12B QAT on llama.cpp 7/13. That last one was faster, but adding Gemma 4's new MTP drafter gave no speed-up on a 6 GB GPU, because when the model is split between GPU and CPU, checking the draft costs as much as it saves. Every model did fine on clean internet photos. Real, messy garden photos are where Gemma pulled ahead.

What still doesn't work

Some plants are still wrong: moringa (called curry leaf; both have small leaflets), a ZZ plant, an Indian almond tree and one rose plant. That's why the app always shows "could also be", Mystery finds and the ✏️ correction.
Birds and insects: the field guide doesn't cover them yet. My one bird test is still a miss.
Its confidence is a hint, not a measurement. It says 85% for right and wrong answers alike.
My test labels are careful judgements, not a botanist's. A couple of photos I left out of the score because I couldn't identify them myself.

I built this with Claude Code as my pair programmer, including the scoring scripts that made every decision above a measured one.

Why Does Open Innovation Matter?

For this project, open isn't a nice-to-have. Every fix above depended on it.

It works where there's no signal. Gardens and parks are where mobile data is weakest. The phone only needs a local link to the laptop.
My photos stay mine. Photos of my street and garden, with times attached, are location data. They never leave my phone and laptop.
It costs nothing per photo. I ran hundreds of test photos through Gemma while building this, with no API bill and no rate limit. A closed API would have made that kind of measuring expensive.
I could look inside and choose the fix. With an open model I could try prompts, swap models, quantize, add an MTP drafter, and fine-tune the weights myself, then measure which actually worked. The answer (a text file beat a fine-tune) is only something you can learn when you're allowed to try both.
Anyone can make it local. field_guide.csv is a plain file. Someone in Kerala, Kenya or Kansas can replace it with their own region's plants, and the same open model becomes their field guide. That's the part I'm most excited about: open models plus open, editable local knowledge.

Prize Categories

Best Use of Gemma. Gemma 4 12B (open weights) runs locally through Ollama and does all of the vision identification, the second-look reasoning, the quests and the walk journal. I also fine-tuned it (LoRA on a DGX Spark) and tested its MTP drafter and QAT version along the way.

Now close the laptop and go and see what's growing by your gate. 🌿

We made SigNoz's LLM observability actually reachable…

Gajanand V Dhayagode — Sun, 26 Jul 2026 17:01:23 +0000

-and added the one signal it was missing

Track 01 — AI & Agent Observability. A hackathon project built from scratch, July 20–26, 2026.

There's a sentence we said to each other on day one and then kept saying all week, because it kept being true:

An API 200 is not evidence that a feature works.

This post is the story of a project that lives entirely inside that sentence. It's about a feature SigNoz already shipped that almost nobody can reach, a real bug that makes part of it impossible, and a missing signal — measured answer quality — that we built on top. And it's about how we used SigNoz not as a black box but as the thing under test.

The surprising starting point: SigNoz already built the machinery

Most "add observability to LLMs" projects start from zero. Ours didn't, and that's the interesting part.

SigNoz v0.133.0 already ships a complete LLM-observability product in open source: a dedicated frontend module, two custom OpenTelemetry collector processors (an LLM-pricing processor and a span-mapper processor), and CRUD APIs behind them. The machinery to ingest gen_ai.* spans, price them, normalize vendor-specific shapes, and render them on a dashboard is there, in the OSS repo, today.

So our thesis wasn't "SigNoz is missing LLM observability." It was sharper and, we think, more useful:

SigNoz shipped the machinery. We make the shipped feature actually reachable — and add the one signal it's missing: measured answer quality.

That framing mattered to us ethically, too. Nothing in this project reads "SigNoz is broken." It reads: they built something good, three defaults and one bug stand between a self-hoster and it, and here's the fix — upstreamed.

The three silent gates

Here's what actually stands between a fresh docker compose up and a working LLM-observability feature. Every one of them fails silently — the backend returns success while the frontend hides the result, and nothing tells you why.

Gate 1 — `enable_ai_observability` is default-disabled

The whole feature sits behind an experimental flag that ships off. With it off, the UI route for LLM observability silently redirects to /home. No error, no "coming soon," no toast. You click the thing and you're just... somewhere else. At the time we hit it, there were zero web results for the flag name.

Gate 2 — `use_dashboard_v2` is default-disabled

The seeded LLM dashboards are v2 dashboards. Import one over the API with this flag off and it succeeds — 200 OK — and then renders as an empty "Welcome to your new dashboard" shell. The data is flowing. The panels are defined. The screen is blank.

This is where the day-one sentence was born. Both gates let the backend succeed while the frontend hides the result. We got bitten by both, in one afternoon, before we understood the pattern.

The fix for gates 1 and 2 is to ship both flags together — and it has to be both:

SIGNOZ_FLAGGER_CONFIG_BOOLEAN_ENABLE__AI__OBSERVABILITY=true
SIGNOZ_FLAGGER_CONFIG_BOOLEAN_USE__DASHBOARD__V2=true

Setting only one still gives you an empty shell — the exact failure we're fixing. So any "dashboards-as-code" deliverable that sets one flag is quietly reproducing the bug.

Gate 3 — the registries ship empty

Two registries back the feature — llm_pricing_rules (model prices) and span_mapper_groups (vendor→spec mapping presets). Both ship with zero rows. No seeded prices, no mapping presets. And there's a nasty second-order trap here: SigNoz's own CrewAI docs tell users to instrument with OpenInference — which emits zero gen_ai.* attributes — so a user who follows the docs faithfully lands on a seeded dashboard that shows nothing, and has no way to know the mapping layer is what's missing.

The one real defect: cost is impossible regardless of configuration

The three gates are configuration. This one is a bug.

SigNoz's LLM-pricing processor and span-mapper processor are generated into the collector config correctly — we verified the rendered YAML. But neither of them is ever wired into a pipeline:

GenerateCollectorConfigWithLLMPricingProcessor and GenerateCollectorConfigWithSpanMapperProcessor both write only collectorConf["processors"].
Neither touches service.pipelines.traces.processors.

In OpenTelemetry Collector terms, that's dead config: a processor that exists but is in no pipeline never runs. So even with the flag on and a correctly-rendered gpt-4o pricing rule, the effective traces pipeline was still [signozspanmetrics/delta, batch], and ClickHouse showed no cost attributes — not _signoz.gen_ai.total_cost, not _signoz.gen_ai.cost_input / cost_output (those are the real emitted names; there is no gen_ai.usage.cost output attribute).

The good news: the codebase already does this right somewhere. The logs feature wires its generated processor straight into the pipeline. So the fix is small, it follows in-repo prior art, and it's the kind of change a maintainer can verify in two minutes:

enable enable_ai_observability → add a pricing rule → observe the effective traces pipeline is still [signozspanmetrics/delta, batch] → cost attributes never appear.

We filed it upstream as SigNoz/signoz#12268 — a small, focused change that cites the prior-art file and ships a clean reproduction. A teammate's fix, not a hackathon drive-by.

The part we're proudest of: measuring whether the answers are any good

Everything above makes the shipped feature reachable. But there's a signal SigNoz's LLM product genuinely does not have, and it's the one that matters most for agents in production: is the model's answer any good, and did the last release make it worse?

Cost, latency, and token counts tell you the model ran. They say nothing about whether it was right. So we built an eval layer.

Emitting a signal SigNoz has never seen

OpenTelemetry specs an evaluation signal — gen_ai.evaluation.result (name, score, label, explanation). It has zero references in the SigNoz codebase. We emit it: an LLM-as-judge scores an answer, and the score travels into SigNoz over OTLP as spec-shaped telemetry.

Crucially, how you carry it decides whether it's usable. We learned this the hard way and built a probe (probe_eval_carriers.py) that emits the same score three ways side by side to prove which one survives:

Span events are stored in ClickHouse as Array(String) — you cannot filter or aggregate inside them. An eval you can't GROUP BY is not auditable.
A spec-compliant log record and a dedicated evaluation span both land as queryable attributes.

So we carry every score on the two that SigNoz can actually aggregate, never on a span event.

The judge earns its place by benchmark, not reputation

The judge is the load-bearing piece: every downstream claim is only as trustworthy as the score it produces. So we picked it by measuring it against our actual task, and we picked it to run locally on open weights — gpt-oss:20b on Ollama, free and self-hostable.

That last choice is architecture, not thrift. This entire project exists to remove a gate between a self-hoster and a feature. A judge that needed a paid API key would reintroduce exactly that kind of gate — now you need a billing account before you can evaluate anything. A free, self-hostable judge works for the same person we're unblocking.

We benchmarked five candidates. The results were not what "bigger is better" would predict:

model	valid JSON	grounded	fabricated	separation	jitter (temp 0)	p50 latency
gpt-oss:20b	3/3	1.000	0.000	+1.000	0.000	3.16s
gpt-oss:120b	3/3	1.000	0.133	+0.867	0.250	4.66s
nemotron-3-nano:30b	flaky	1.000	0.270	+0.730	0.180	5.38s
deepseek-v4-flash	—	HTTP 403 (not in tier)
gemma4:31b	—	returned non-JSON

The bigger model was the noisier one. On a repeated fabricated input at temperature 0, gpt-oss:120b returned 0.25 / 0.00 / 0.15 — a 0.25 spread — while gpt-oss:20b returned 0.00 every single time. A quality-regression alert lives or dies on that stability: a judge that wobbles ±0.25 makes "quality dropped 12%" indistinguishable from measurement noise, and can name the wrong release as the culprit. We chose the model our alert can actually trust, and it happened to be the smaller, faster, free one.

Every score also carries the two attributes that make it reproducible after the fact:

gen_ai.evaluation.judge.model — which model scored it
judge.prompt.version — which rubric version scored it

So "which judge and which rubric produced this score?" is a GROUP BY, not a shrug.

The alert that names the release that regressed

The payoff is a threshold alert: avg(gen_ai.evaluation.score.value) below 0.7, over a 30-minute window, grouped by service.version. When a release's answer quality slides under the floor, the alert fires and names the release. No anomaly detection — that's enterprise-gated in SigNoz — just a threshold on a signal we made aggregatable.

We measured the full loop end to end against a live instance, not estimated it:

emit at 16:21:43Z → alert transitioned to firing at 16:25:40Z
span-to-alert = 238s (~4 min)

And here's the part that matters for anyone building on this: that time is almost entirely SigNoz's own machinery — the 120s eval_delay plus the 1-minute evaluation interval. The judge adds only ~2.5s per call, and because spans are timestamped when the run starts, even that overlaps the eval_delay rather than stacking on it. The eval layer adds essentially nothing to detection time; SigNoz's evaluation interval is the floor.

For the demo we use a deliberately loud drop — measured mean 0.958 → 0.167 — so the mechanism is legible in one glance. That's a demo choice, not a limitation: the alert fires on the same 0.7 floor whether the slide is a cliff or a subtle crossing.

How we deployed it — and made it reproducible

We ran SigNoz self-hosted in WSL2 (Ubuntu 24.04) on native Docker Engine — not Docker Desktop, whose ClickHouse Keeper is known to segfault. Deployment is via SigNoz Foundry (foundryctl), their "one config, one command" tool: you declare the whole deployment in a casting.yaml, forge expands it into platform manifests plus a checksum-pinned casting.yaml.lock, and cast deploys it.

Because the judges of this hackathon may re-run Foundry to reproduce our deployment, we didn't hand-edit generated files — we put the fix in the source of truth. Our casting.yaml sets both gates as first-class env overrides on the SigNoz service:

spec:
  deployment:
    flavor: compose
    mode: docker
  signoz:
    spec:
      version: v0.133.0
      image: signoz/signoz:v0.133.0
      env:
        SIGNOZ_FLAGGER_CONFIG_BOOLEAN_ENABLE__AI__OBSERVABILITY: "true"
        SIGNOZ_FLAGGER_CONFIG_BOOLEAN_USE__DASHBOARD__V2: "true"

We verified this is the right way to do it: forge merges those overrides into the defaults (it doesn't wipe the DB/store env), the flags render into the generated compose.yaml, and — critically — forge is deterministic: re-running it produces a byte-identical lock. So a judge who clones the repo and runs foundryctl forge && foundryctl cast gets the feature-reachable deployment, not the default empty shell. The casting file is part of the argument, not a checkbox.

What we'd tell the next person

Treat a 200 as the beginning of verification, not the end. Both gates that cost us the most time were backends succeeding while frontends hid the result. Visually confirm every dashboard; never trust the POST.
Pick your LLM judge by measuring it, not by its size or name. The stability of the judge is the stability of your alert. Bigger lost.
Choose carriers your backend can aggregate. A beautiful eval signal stuffed into a span event is invisible to GROUP BY. Know your storage model before you emit.
Put the fix in the source of truth, not the generated output — so reproduction reproduces the fixed thing.

SigNoz built a genuinely good LLM-observability product. We made it reachable, filed the bug that was blocking cost, and added the measured-quality signal it was missing — all in OSS, all reproducible from one casting.yaml.

Repo (with casting.yaml + casting.yaml.lock for reproduction): https://github.com/gajanand27-05/SigNoz-conform