Syapa: The Sloth's Secret Weapon Against Freelance Scope Creep (Powered by Gemma & Local AI)

mannubaveja007 — Sun, 25 Oct 2026 17:13:28 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady

What I Built

Every freelancer knows the dread of post-meeting administrative chores.

You finish an intense 60-minute call with a client. Instead of relaxing, you face an exhausting to-do list: re-listening to call recordings, combing through pages of notes, spotting where the client casually snuck in unbilled extra features ("Can we also just add an SMS notification drip and student portal?"), catching moments where you accidentally gave away free work ("Sure, I can look into that this weekend"), and drafting an uncomfortable pushback email.

In honor of International Sloth Day, I built Syapa (meaning “chaos” or “headache” in Punjabi/Hindi)—a precision freelance intelligence platform designed to do all of that heavy lifting so you don't have to.

Syapa lets freelancers channel their inner sloth:

One-Click Audio Capture: Click "Start Recording" inside a lightweight Chrome Manifest V3 side panel during any Google Meet, Zoom, or browser call.
Zero Post-Meeting Effort: Once the call ends, you literally walk away. Syapa’s asynchronous background pipeline automatically transcribes the conversation with speaker diarization ("You" vs. "Others"), chunks and embeds the dialogue locally, and runs Google Gemma to extract an executive summary, action items, sneaky scope creep, and dangerous unbilled promises.
Deterministic Hallucination Filter: All extracted quotes are verified verbatim against the raw transcript using deterministic string validation. No fabricated client statements or hallucinated promises.
Automated Boundary Coaching: Syapa generates copy-pasteable negotiation pushbacks and formal change orders ready to send to the client.
Interactive Sparring Room: Before having that difficult conversation, practice your pushback against realistic client personas (like Mr. Sharma, the budget-pinching coaching center owner) using conversational AI.

Demo

🎥 Video Walkthrough & Architecture Tour:

[

(Local-First Architecture: Because Syapa runs local edge inference via Ollama and a native Chrome Manifest V3 side panel for privacy reasons, the video walkthrough demonstrates the full end-to-end recording, pipeline processing, and interactive sparring session live).

Code

Syapa is completely open-source under the MIT License:

mannubaveja007 / Syapa

Syapa 🛡️

A precision freelance intelligence platform combining a Chrome Manifest V3 meeting recorder with server-side speaker diarization, grounded negotiation coaching, and real-time Conversational AI client sparring.

Architecture Overview

                                      +---------------------------------------------+
                                      |            Browser (Chrome MV3)             |
                                      +---------------------------------------------+
                                      |                                             |
                                      |   [Active Tab Audio]   [Microphone Audio]   |
                                      |           |                    |            |
                                      |           +---------+----------+            |
                                      |                     |                       |
                                      |                     v                       |
                                      |       AudioContext Stream Mixer             |
                                      |       (Tab Audio routed to speakers)        |
                                      |                     |                       |
                                      |                     v                       |
                                      |     MediaRecorder (audio/webm;codecs=opus)  |
                                      |                     |                       |
                                      +---------------------|-----------------------+
                                                            |
                                                            | POST /api/meetings/transcribe
                                                            | (x-api-token: syp_...)
                                                            v
+-----------------------------------------------------------------------------------+
|                            Syapa Web App (Next.js 14)                             |
|                               http://localhost:3000                               |
+-----------------------------------------------------------------------------------+
|  POST /api/meetings/transcribe   --> Validates audio, calls ElevenLabs STT        |
|                                      Server-side only (diarize=true, scribe_v2)   |
|                                      Rejects empty/silent audio (HTTP 422)        |
|                                      Saves TranscriptSegments (You / Others)      |
|                                      Triggers asynchronous post-meeting pipeline  |
|  GET  /api/meetings/:id/analysis --> Polls status, executive summary & top

…

View on GitHub

Architecture Overview

                                  +---------------------------------------------+
                                  |            Browser (Chrome MV3)             |
                                  +---------------------------------------------+
                                  |   [Active Tab Audio]   [Microphone Audio]   |
                                  |           \                    /            |
                                  |            AudioContext Stream Mixer        |
                                  |                       |                     |
                                  |                       v                     |
                                  |             MediaRecorder (Opus)            |
                                  +-----------------------|---------------------+
                                                          | POST /api/meetings/transcribe
                                                          v
+---------------------------------------------------------------------------------------+
|                              Syapa Web App (Next.js 14)                               |
+---------------------------------------------------------------------------------------+
| - Speaker Diarization: ElevenLabs Scribe v2 (Server-side key protection)              |
| - Embeddings Engine: nomic-embed-text (Fixed 1024-dim, single batched request)        |
| - Reasoning & Extraction: Gemma 4 / Gemma 2 via Ollama local inference                |
| - Hybrid RAG: 70% pgvector HNSW cosine + 30% GIN tsvector full-text search            |
| - Deterministic Verification: Verbatim string quote matcher                           |
| - Client Sparring: ElevenLabs Conversational AI + Gemma client state simulator        |
+---------------------------------------------------------------------------------------+

How I Built It

Syapa is structured as a performant TypeScript monorepo (pnpm workspaces) combining modern web tech with local open-source AI infrastructure:

1. Open-Source AI Stack at the Core

Google Gemma (gemma4:12b / gemma-2-9b-it) via Ollama: Gemma powers the core intelligence pipeline. Using strict JSON schema generation, Gemma inspects the dialogue, analyzes cross-meeting behavioral patterns against past client calls, flags high-risk freelancer concessions, and drafts assertive counter-proposals.
nomic-embed-text Local Embeddings: Transcripts are chunked into 350-word sliding windows and processed in one batched HTTP request using nomic-embed-text running on localhost:11434. Vectors are normalized to 1024 dimensions and stored in PostgreSQL.
PostgreSQL + pgvector Hybrid Search: Client briefs and past transcripts are indexed using an hnsw ("embedding" vector_cosine_ops) index alongside PostgreSQL tsvector indexes. When practicing negotiations, Syapa pulls relevant context using a 70% vector + 30% lexical full-text hybrid query with sub-10ms latency.

2. Plain Chrome MV3 Extension (`apps/extension`)

Loads in < 5 ms with zero build or bundle overhead.
Uses navigator.mediaDevices.getDisplayMedia and getUserMedia with an in-memory AudioContext mixer. Tab audio is piped to local speakers so you can hear your client without audio feedback loops.
Empty or silent recordings (<1000 bytes) are rejected automatically at the gateway before hitting any models.

3. Deterministic Quote Verification (`packages/ai/src/quotes.ts`)

LLMs often paraphrase when asked to quote what someone said. For freelance disputes, an ungrounded quote is fatal. Syapa runs a deterministic string-matching verification step in plain code after Gemma completes its analysis. If a quote cannot be verified 100% verbatim against the transcript, it is automatically discarded and logged.

4. Conversational AI Practice Room (`apps/web/src/app/practice`)

Integrated with the @elevenlabs/react SDK, freelancers can vocally spar against realistic client personas like Mr. Sharma:

Runs a Class 9–12 coaching center.
Stated budget: ₹12,000 | Absolute ceiling: ₹15,000.
Hidden asks: Student fee payment gateway, PDF login portal, and Hindi localization.
Freelancers practice holding their price ground and setting scope boundaries before the real call.

Why Does Open Innovation Matter?

Why not just hook this up to a proprietary closed API like OpenAI or Anthropic? For freelance intelligence, open innovation is not just a preference—it is a functional necessity:

Client Confidentiality & NDA Compliance: Freelancers routinely discuss proprietary codebases, unreleased product roadmaps, and sensitive budget numbers on client calls. Routing unredacted audio or transcripts to closed cloud models violates standard client NDAs. By running Gemma and nomic-embed-text locally through Ollama, client data never leaves the freelancer’s machine.
Zero Marginal Cost for Heavy Workloads: Transcripts and audio embeddings generate hundreds of thousands of tokens over the course of a multi-week project. Local open-weight models let freelancers run extensive cross-meeting pattern detection, RAG queries, and simulation sessions all day long for $0.00.
No Vendor Lock-In or Deprecation Surprises: Closed APIs regularly deprecate models, shift system prompt behaviors, or adjust safety guardrails that break structured output schemas. Open-weight models like Gemma give the developer complete control over execution, temperature, sampling, and reproducibility.

The Lazy Payoff

Here is what channeling your inner sloth looks like in practice with Syapa:

Traditional Freelancer Routine	The Syapa "Sloth" Way
45 mins re-listening to the call & taking manual notes	0 mins: Syapa automatically diarizes and summarizes the call
30 mins combing through emails and SOWs to check if a feature was included	Instant: Gemma highlights scope creep with verified verbatim quotes
20 mins stressing over how to politely say "that will cost extra"	1-click: Copy the pre-drafted boundary email and change order proposal
Unbilled hours lost to silent concessions you forgot you made	0 hours: Risky promises are flagged immediately with risk analysis

Total Time Saved: ~1.5 to 2 hours per client meeting, plus thousands of dollars saved in prevented scope creep.

Prize Categories

🌟 Best Use of Gemma ($200 Featured Category)

Syapa relies heavily on Google Gemma (gemma4:12b / gemma-2-9b-it) as its reasoning backbone:

Gemma Structured Analysis: Executes strict schema-enforced extraction of meeting agreements, action items, and risky commitments (apps/web/src/lib/pipeline.ts).
Cross-Meeting Behavioral Pattern Analysis: Gemma compares the current transcript against past completed client calls to detect recurring concession patterns.
Dynamic Negotiation Simulator: Gemma powers the "Bad Client" AI opponent, dynamically managing hidden emotional states, budget thresholds, and concession triggers during sparring sessions (apps/web/src/app/api/simulator/[id]/message/route.ts).

🎙️ Best Use of ElevenLabs ($100 Partner Category)

Syapa integrates ElevenLabs across two distinct modalities:

ElevenLabs Scribe v2 Speech-to-Text: Server-side audio transcription with word-level speaker diarization, securely mapping audio into "You" vs. "Others" dialogue blocks without leaking API keys to the browser extension.
ElevenLabs Conversational AI SDK (@elevenlabs/react): Powers low-latency real-time voice sparring with authentic client archetypes in the Practice Room.

Built with passion (and maximum conservation of energy 🦥) for Hacktoberfest 2026.

My Uncle’s Khata Notebook Was Failing Him --> So I Built Hisabb

mannubaveja007 — Sun, 04 Oct 2026 20:32:20 +0000

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

Hisabb — बोलो. हिसाब हो गया.

I built Hisabb for my uncle, who runs a grocery shop and manages dozens of customers buying on udhaar—credit—every day.

He records everything in a physical notebook: customer names, groceries purchased, amounts owed, and payments received. But as transactions accumulate, finding a customer’s history means flipping through pages, calculating balances by hand, and checking whether a payment was already recorded.

Sometimes entries are unclear, missed, or overwritten. That creates more than extra work—it can cause confusion between my uncle and his customers.

Hisabb brings that notebook into one simple digital ledger. It helps him track each customer’s udhaar, payments, transaction history, and outstanding balance without searching through pages.

Its voice-first workflow makes adding entries easier: speak a transaction in Hindi or Hinglish, review the extracted details, correct anything the speech recognition misunderstood, and save. Manual entry is available too.

The goal is simple: less time managing hisab, fewer mistakes, and more trust between the shopkeeper and the customer.

Demo

Start with our visual introduction:
Visit Hisabb’s welcome page best animation works are here!

Explore the working ledger:
Open Hisabb

The /welcome page introduces Hisabb’s story and purpose; the ledger shows how that idea becomes a practical bookkeeping workflow.

DEV Community: mannubaveja007

Syapa: The Sloth's Secret Weapon Against Freelance Scope Creep (Powered by Gemma & Local AI)

What I Built

Demo

Code

mannubaveja007 / Syapa

Syapa 🛡️

Architecture Overview

Architecture Overview

How I Built It

1. Open-Source AI Stack at the Core

2. Plain Chrome MV3 Extension (apps/extension)

3. Deterministic Quote Verification (packages/ai/src/quotes.ts)

4. Conversational AI Practice Room (apps/web/src/app/practice)

Why Does Open Innovation Matter?

The Lazy Payoff

Prize Categories

🌟 Best Use of Gemma ($200 Featured Category)

🎙️ Best Use of ElevenLabs ($100 Partner Category)

My Uncle’s Khata Notebook Was Failing Him --> So I Built Hisabb

What I Built

Demo

2. Plain Chrome MV3 Extension (`apps/extension`)

3. Deterministic Quote Verification (`packages/ai/src/quotes.ts`)

4. Conversational AI Practice Room (`apps/web/src/app/practice`)