DEV Community: Tenath Dilusha

Echoes of Lanka - A Sri Lankan historical roleplay altar powered by open-weight AI

Tenath Dilusha — Fri, 30 Oct 2026 13:48:59 +0000

What I Built

Echoes of Lanka is a Sri Lankan historical roleplay altar powered by open-weight AI.

You sit with imagined voices from the island's past — King Kashyapa on Sigiriya, anonymous fresco painters, Anuradhapura stupa masons, Sanghamittā Therī, Parākramabāhu I, and coastal cinnamon traders — while verified claims stay visibly separate from fiction.

Week 4 asked for something spooky or something that honors the past in the spirit of Día de los Muertos. I chose remembrance over jump-scares: when the line between past and present feels thin, we should still know which side of the altar is candlelight and which is stone.

So every reply splits into two panels:

Verified Echoes — curated facts retrieved without an LLM (keyword scoring over a small corpus, tagged chronicle / archaeology / scholarship, with sources)
Living Voice — Gemma roleplay, explicitly labeled as imagined dialogue, constrained to those facts

The LLM never decides what is true. It only speaks beside the record.

Voices you can meet

Echo	Who they are
King Kashyapa I	Lord of the Lion Rock, Sigiriya
The Fresco Painter	Anonymous hand of the Mirror Wall
A Stupa Mason	Brick courses of Anuradhapura
Sanghamittā Therī	Bearer of the Bodhi sapling
Parākramabāhu I	The great irrigator-king
A Cinnamon Trader	Coastal exchange and spice routes

Demo

Repo: https://github.com/TenathDilusha/Echoes-of-Lanka

npm install
ollama serve          # if not already running
ollama pull gemma3:4b # or gemma3:1b on smaller machines
cp .env.example .env.local   # optional
npm run dev

Open http://localhost:3000, pick a character, ask something like "Why did you build on the rock?" or "How did you mix your colors?"

You'll see:

Curated claims + confidence tags + sources on one side
First-person atmospheric dialogue on the other — clearly marked as imagined
A status chip for whether Ollama/Gemma is reachable; if the model is down, a fact-grounded fallback voice still answers so the altar never goes fully silent

Code

MIT-licensed Next.js app. The open pieces that make it work:

Piece	Role
Gemma (`gemma3:4b` via Ollama)	Living Voice — local open-weight chat
Deterministic retrieval	Verified Echoes — no embeddings, no LLM ranking
System prompt rules	Dialogue may only lean on retrieved facts; gaps stay gaps
Fallback template	Offline / Ollama-down path still grounded in facts

The critical design choice is in the system prompt: the model is told its reply is imagined dialogue only, and concrete historical claims must come from the retrieved fact block — or it must admit the chronicles are silent.

CRITICAL RULES — open historical honesty:
1. Your reply is IMAGINED DIALOGUE only — a creative echo, never a primary source.
2. You may ONLY lean on the VERIFIED FACTS listed below for concrete historical claims.
3. If asked something outside those facts, say you do not remember / the chronicles are silent
   — do not fabricate battles, dates, or monuments.

Retrieval is intentionally boring and auditable:

/** Deterministic keyword retrieval — no LLM involved in FACT selection. */
export function retrieveFacts(characterId, query, limit = 4) { /* ... */ }

That split is the whole product thesis: open models for presence; open, inspectable data for truthfulness.

How I Built It

Stack

Next.js + TypeScript UI (character grid → chat)
Ollama serving Google Gemma 3 locally (OLLAMA_URL / OLLAMA_MODEL)
A curated fact bank with confidence labels and source lines
Keyword scoring so fact selection is reproducible and cheap

Architecture

User question
    │
    ├─► retrieveFacts()  ── deterministic score over curated corpus
    │         │
    │         └─► VERIFIED ECHOES (claims + sources + confidence)
    │
    └─► Ollama / Gemma  ── roleplay only, constrained to retrieved facts
              │
              └─► LIVING VOICE (explicitly imagined)

Theme interpretation

Día de los Muertos is about ofrendas — offerings that keep memory warm without confusing the living for the dead. Echoes of Lanka treats historical AI the same way: the Living Voice is the marigold and candle; the Verified Echoes panel is the photograph and the name card. You can feel the presence and still read the label.

Halloween's "ghost in the machine" angle still fits: the model is a simulation that refuses to pretend it's possession or a primary source.

Why Open Innovation Matters

National memory and family-adjacent stories should not require shipping every prompt to a closed API.

Open worked better here than closed would have:

Runs on a laptop — once Gemma is pulled, chat does not need a cloud key. Cost to run the dialogue path is electricity + hardware you already own.
Works when the network is gone — Ollama local; fallback voice if the daemon is down. Cultural exploration does not stop at a captive portal.
You can swap the model — OLLAMA_MODEL=gemma3:1b (or another Ollama tag) without rewriting the product. Closed APIs lock voice, pricing, and availability together.
Facts stay under human control — retrieval is deterministic code over a curated corpus. A closed RAG stack would still let the model blur "source" and "story"; we refused that blur on purpose.
Nothing culturally sensitive has to leave the machine — educational roleplay about heritage should not depend on a vendor that might change terms, disappear, or train on every whispered question.

Community builders have been saying the same thing about local open-weight tooling for privacy and offline use — for example Building a Privacy-First Local AI Assistant and Day 16: Local Model Execution with Ollama. Echoes of Lanka applies that lesson to historical honesty, not only to secrets and code.

The Moment It Landed

The first time I asked the fresco painter "Who were the women you painted?" and watched Verified Echoes light up with scholarship-tagged claims while the Living Voice answered in pigment and plaster — without inventing a royal romance the corpus did not support — that was the ofrenda click.

The second time: I stopped Ollama on purpose. The altar still answered from the fact bank and admitted the mist. That mattered more to me than a prettier hallucination.

This is an educational simulation. Chronicle narratives are literary as well as historical. Imagined dialogue is fiction grounded by a small curated fact bank — not a substitute for academic research or living cultural authority.

I Built My Friend a Private Coach That Says What to Fix Before Applying to Jobs. No more applying, waiting, silence.

Tenath Dilusha — Mon, 05 Oct 2026 04:46:44 +0000

What I Built

If you are a university student, you probably know this loop:

The problem isn't effort. It's feedback. Nobody tells you which skills you're missing, which project you should lead with, or what the interviewer will ask.

The obvious fix is to paste your CV into an AI chatbot. But a CV holds your name, phone number, home address, university and GPA. My friend didn't want to send that to a server they can't see inside. Honestly, neither would I.

The solution: JobPilot 🧭

JobPilot is a private AI career coach that runs entirely on your own laptop. You give it your CV and a job you want, and it tells you exactly what to fix before you apply.

Your CV never leaves your laptop. The AI is Gemma 3, running locally with Ollama.
Free. No subscription, no API bill, no sign-up.
Works offline. Everything except live job search works with Wi-Fi off.
Honest. It doesn't invent skills, experience or fake numbers, and the code checks that (more on that below).
Made for Sri Lankan students. It understands internships, trainee roles, GPA, A/L results, and keeps the English simple.

To show what it does, I'll use a sample student, Nimali: a Computer Science undergrad at UCSC with three projects (a React club website, a paddy price predictor in Python, and a Java library system). She's applying for a Software Engineering Intern role in Colombo.

1. 🎯 Match & gaps: "How close am I?"

One click compares her CV with the job.

First, an instant skills match appears. No AI here, just a fast keyword scan: she has 4 of the 17 skills the job lists (JavaScript, SQL, React, Git).

Then Gemma reads both documents properly and gives:

An overall fit score (65%) and a two-line summary

Her strengths, with proof copied word for word from her CV, like JavaScript · "Languages: Python, Java, JavaScript, SQL"

Missing skills ranked by importance, each with a concrete way to learn it. Top of the list: Node.js & Express.js (HIGH): "Build a simple API using Node.js and Express.js with a small dataset."

CV edits that only use things she actually did

Five interview questions tailored to her projects and this job

Match and gaps: keyword match, AI fit score, strengths with CV quotes, ranked skill gaps and interview questions

2. 🎤 Interview practice: "Would that answer land?"

Click Practice on any question, type your answer the way you'd say it, and JobPilot plays the interviewer.

Nimali answered a question about her Library Management System with "...The hard part was the tables for borrowing, but we fixed it." JobPilot gave her a 4/10, and told her exactly why:
"Needs more detail: what tables did you design? What were the key fields?"
"Avoid vague phrasing like 'we fixed it'."

Then it wrote a stronger answer in her own voice, using only her CV and what she said.

Interview practice: tailored questions, the student's answer, a 4/10 score, what worked, how to improve and a stronger answer

3. ✍️ Rewrite: from "made a website" to a CV bullet

Most students describe projects like this:

Made a website for our university club where members can register for events and get reminders. Used React and Firebase.

JobPilot turns it into a polished description and CV bullets like "Utilized Firebase to manage event registrations and send automated reminders to club members."

The important part is what it doesn't do. It never makes up numbers. Instead, it asks the student to fill in the truth: "Quantify the impact, for example 'Supported [X] club members'."

Rewrite: a one-line project description turned into a polished paragraph, CV bullets and suggestions

4. 💬 Ask JobPilot: a coach who has read your CV

A chat that already knows your CV and the job. The answer streams in word by word, so you're not staring at a spinner.

I asked "Which of my projects should I highlight for this job?" and it picked Campus Event Hub (the React one, matching the job's stack) and explained why the other two are weaker fits for this particular role.

Ask JobPilot chat recommending the Campus Event Hub project for a React internship

5. 📈 Find jobs: "What should I learn next?"

This is the feature I'm proudest of. Search for real jobs (live from Google Jobs via SerpApi), and JobPilot:

Ranks every posting by how well it matches your CV, and shows what's missing for each one.

Counts which skills employers are asking for across all the postings. Blue bars are on your CV; yellow bars aren't yet.

Asks Gemma to turn that chart into a learning plan: the most-requested skills you're missing, each with a first step you can do this weekend.

When I searched "software engineering intern" in Sri Lanka, it pulled 17 real postings (Softmint, LSEG, Fidenz, Codimite and more). The results were eye-opening:

Problem Solving was asked for in 10 of 17 postings, more than any programming language.

JavaScript and React were in 8 of 17.

Gemma's plan for Nimali: learn OOP (6 of 17), then Node.js (5 of 17), then Agile (4 of 17), and put Python, Java and SQL at the top of her CV.

That's the kind of advice a senior would give you, except it's based on this week's real job posts.

Find jobs: employer skill demand chart across 17 live postings, a Gemma learning plan, and job cards ranked by CV match

Code

TenathDilusha / jobpilot

Private AI career coach for Sri Lankan students. Compares your CV with a job, finds skill gaps, and preps you for interviews, using Gemma running locally.

JobPilot: Career Assistant for Sri Lankan Students

JobPilot is a private AI career coach for Sri Lankan university students applying for internships and entry-level jobs. Upload your CV, paste a job description, and ask:

"What skills am I missing for this internship?"
"Rewrite my project description."
"What questions might they ask me?"
"Compare my CV with this job description."

Everything runs on your own laptop with an open-weight Gemma model served by Ollama Your CV is never uploaded to a cloud AI service.

Features

Tab	What it does
Match & gaps	Instant keyword skill match, then Gemma's analysis: fit score, strengths with evidence quoted from the CV, missing skills ranked by importance with how to learn each, CV edits, and tailored interview questions
Rewrite	Turns a rough project description into a polished paragraph and action-verb CV bullets, using `[X%]` placeholders instead of invented numbers
Interview practice	Answer a question

…

View on GitHub

MIT licensed. Python + FastAPI backend, plain HTML/CSS/JavaScript frontend, 29 tests. Setup is three commands (see the README).

How I Built It

Here's the whole system on one page:

JobPilot architecture: CV and job description go into the laptop, through text extraction and a skill scanner, into Gemma 3 in Ollama, out to five features. SerpApi only receives the search words.

Backend: FastAPI. It reads PDFs (pypdf) and Word files (python-docx), and exposes one endpoint per feature.

AI: Gemma 3 through Ollama's /api/chat, with JSON-schema structured outputs for every analysis and token streaming for chat.

Skill scanner: 100 skills with aliases (ReactJS, React.js and react all count, but Java never matches JavaScript).

Job search: SerpApi's Google Jobs engine, two pages of results, cached for 6 hours so free search credits last.

Frontend: no framework. Every piece of model output is inserted as text, never as HTML.

Lesson 1: a small model needs guardrails, not just instructions 🛡️

Gemma 3 4B is impressive for its size, but small models love to make things up. In my early tests, it praised Nimali for "maintaining a website using Git" (her CV never says that) and suggested adding "Developed REST APIs with Node.js" to her Python project.

Telling the model "never invent things" in the prompt helped, but it still ignored the rule sometimes. So I stopped trusting the prompt alone and built five layers:

Keeping a 4B model honest: read the documents, scan skills without AI, prompt Gemma with rules, force a JSON schema, then check the answer in code

The last step is a few lines of plain Python. A "strength" only survives if its quote really appears in the CV, and any CV tip or rewritten bullet that names a skill you don't have gets dropped. The model can be creative; the code makes sure it's never dishonest.

Lesson 2: design for a slow laptop 🐢

I built and tested this on my own laptop: 8 GB of RAM, no GPU. Gemma 3 4B writes about 3-4 words per second there, so every design choice was about making waiting feel OK:

The skills match is instant (no AI), so you see something useful in under a second.

Analysis and interview questions are separate, smaller calls that appear one after another, each with a live timer.

Chat streams word by word.

Outputs are deliberately short. Every extra word costs time.

A ⚡ Fast mode switch swaps to Gemma 3 1B when you're in a hurry.

A cloud model would be faster. But for something you do once per application, a few minutes is a fair price for keeping your CV private.

Why Open Matters

A typical AI CV tool sends your CV to someone else's server. JobPilot keeps it on your laptop, and SerpApi only sees the search words.

This project only works because Gemma is open:

Privacy by design, not by promise. The model runs on the student's own machine, so there's no "we don't store your data" policy to trust. The data just never leaves.

Free forever. Students can check every single application without paying per request.

Offline. Patchy campus Wi-Fi or no mobile data? Everything except job search still works.

Swap the model. 4B for quality, 1B for speed, or point it at a bigger Gemma on a stronger machine with one setting.

Even the job search is private. The only thing sent to SerpApi is the search words, like "software engineering intern, Sri Lanka". Matching postings against your CV happens locally.

Built with ❤️ for a friend during Hacktoberfest 2026. I'm a 3rd-year Computer Science and Engineering student at the University of Moratuwa. If you're job hunting too, give JobPilot a try and tell me what it caught on your CV.

First post

Tenath Dilusha — Sun, 04 Oct 2026 05:04:15 +0000

DEV Community: Tenath Dilusha

Echoes of Lanka - A Sri Lankan historical roleplay altar powered by open-weight AI

What I Built

Voices you can meet

Demo

Code

How I Built It

Stack

Architecture

Theme interpretation

Why Open Innovation Matters

The Moment It Landed

Links

I Built My Friend a Private Coach That Says What to Fix Before Applying to Jobs. No more applying, waiting, silence.

What I Built

The solution: JobPilot 🧭

1. 🎯 Match & gaps: "How close am I?"

2. 🎤 Interview practice: "Would that answer land?"

3. ✍️ Rewrite: from "made a website" to a CV bullet

4. 💬 Ask JobPilot: a coach who has read your CV

5. 📈 Find jobs: "What should I learn next?"

Code

TenathDilusha / jobpilot

Private AI career coach for Sri Lankan students. Compares your CV with a job, finds skill gaps, and preps you for interviews, using Gemma running locally.

JobPilot: Career Assistant for Sri Lankan Students

Features

How I Built It

Lesson 1: a small model needs guardrails, not just instructions 🛡️

Lesson 2: design for a slow laptop 🐢

Why Open Matters

First post