# Samrat Mukherjee — projects > Shipped products and what makes each one work. AI Engineer · AI Agent Developer. ## Numbers - **88%** — autonomous answer coverage. A RAG rebuild of the answer layer at Shortlistapp raised how often the agent answers an employer's open-ended question on its own. - **28%** — lower LLM cost. Same rebuild — retrieving facts instead of re-prompting a large model per question. - **749** — tests on Stockpot. Including an adversarial suite against the agent's tool boundary. - **~15×** — fewer Redis commands. 2.6M → 173K per month, by replacing a polling scheduler with a Lua timer wheel. ## Projects ### Shortlistapp — An agent that applies to jobs end to end. 2026 · Live · paying customers · https://shortlistapp.co It reads the posting, fills the application, answers the employer's open-ended questions from the user's real background, and submits on real ATS portals. Headline metric: **88%** autonomous answer coverage - Stateful LangGraph workflow that resolves every field through tool calls and escalates rather than guessing. - RAG over each user's resume and history so answers come from real background — 88% autonomous answer coverage, 28% lower LLM cost. - A FactGuard layer blocks the model on any field with no retrieved fact, so it never fabricates on a submission the user cannot take back. - Idempotency keys and crash-safe replay guarantee no duplicate applications. Stack: Python, LangGraph, Pinecone, FastAPI, OpenTelemetry ### Icebreakr — AI outreach that researches a person before it writes. 2026 · Live on Chrome Web Store · https://useicebreakr.com Drafts personalized connection requests, DMs and cold emails using LLMs, behind a provider architecture with automatic fallback chains. Headline metric: **4-axis** LLM-as-judge eval harness - Plug-and-play provider architecture — Claude, OpenAI and Gemini behind interfaces with automatic fallback chains, swappable from a single config line. - Hardened against prompt injection: untrusted data isolated in delimiters so the model treats it as data, never instructions, with a typed-error retry loop. - Eval harness with deterministic auto-checks plus an LLM-as-judge scoring four subjective axes, using baseline/diff modes to catch prompt regressions before merge. - Go backend (Gin, PostgreSQL) with errgroup-orchestrated parallel fetches behind a 75s pipeline timeout, per-IP and per-user rate limiting, and daily usage caps. - Manifest V3 Chrome extension in TypeScript using a closed Shadow DOM overlay, with WorkOS OAuth/SSO and a signed handoff that never exposes tokens to JavaScript. Stack: Go, LLM, Next.js, Chrome MV3, TypeScript ### Stockpot — Agents that trade surplus food between restaurants. 2026 · Agentic Commerce Hackathon · https://getstockpot.shop Settling on real payment rails, where money moves only after delivery is verified. Headline metric: **749** tests, incl. adversarial suite - A hard LLM boundary: the matching engine is fully deterministic and agent tools accept no model-supplied numbers. The model chooses which check to run; Python computes every value. - Idempotent settlement protocol keyed on (listing, delivery) — the charge row is written pending before any network call, so a provider timeout retries instead of silently reporting success. - Card rails with network-enforced spending caps and partial capture: the buyer approves a ceiling once by passkey, and a charge above it is refused by the network. - Cut Redis command volume ~15× (2.6M → 173K/month) by replacing a polling scheduler with a Lua timer wheel that sleeps until the next deadline. Stack: Python, FastAPI, PostgreSQL, Redis Source: https://github.com/0xSamrat/stockpot ### NetagiriFiles — Civic transparency for ~800 Lok Sabha MPs. 2026 · 500+ monthly visitors · https://netagirifiles.fun Surfaces self-declared affidavit data with interactive D3.js charts, an India choropleth and per-MP profile pages. Headline metric: **<10 min** full dataset refresh - Concurrent Go ingestion pipeline (goroutines, token-bucket limiter) writing to Postgres via idempotent upserts, refreshing the full dataset in under 10 minutes. - Frontend in Next.js, React and Tailwind with per-MP profile pages indexed for search, serving 500+ monthly visitors. Stack: Go, Next.js, PostgreSQL, D3.js ## Hackathons & achievements - **Soonami Venturethon (Cohort 5)** — Won a $1,000 grant for Defipe.io; grew it to 500+ testnet users and a 1,000+ member Discord. (https://app.foundance.org/projects/11247) - **ETHGlobal — Connext Pool Prize winner** — Cross-chain DAO governance with multichain proposal and voting execution. (https://ethglobal.com/showcase/ape-dao-2-0-i846k) - **ETH for ALL — Arcana Pool Prize winner** — Decentralized token swap via 0x Aggregator with social-login wallet onboarding. (https://devfolio.co/projects/defipe-b5ba)