I'm Animesh — an applied AI/ML engineer with 1 year 9 months shipping LLM-powered products, RAG pipelines, and real-time agentic systems at Formee Holdings. IIT Bombay graduate coursework in CS (CPI 8.72). Daily driver of Claude Code for end-to-end agentic development.
Gateway to graduate research at IIT Bombay.
A WebRTC + Asterisk + Twilio voice/chat CRM with LlamaIndex-driven live response suggestions over active conversation transcripts and DB lookups. Agent desktop, dialplan, and post-call agentic lead analysis — end-to-end.
A one-wallet router across fal.ai's full catalog (Kling, Seedance, Wan, Flux, Veo + 12 more). Six generation modes routed through one dashboard, async queue + webhooks, Director Mode for script→storyboard→parallel renders→stitched MP4. ~18.7k LOC, solo end-to-end.
Full fine-tuning loop on a homelab RTX 4060 Ti (Unsloth, 4-bit NF4). SFT on Dolly-15k, DPO on UltraFeedback. Evaluated with held-out perplexity, Claude Sonnet as pairwise LLM-judge with A/B swap, mini-IFEval, and weight-delta diagnostics.
Homelab control panel — FastAPI + Next.js 16 / React 19. System monitoring, Docker container management with live logs, Jellyfin media integration, mood-based movie discovery, file explorer, plus a TipTap-backed Notes + Tasks workspace. 6+ external systems behind a typed adapter layer.
FastAPI + HTMX + Jinja2 interview-prep tool. Voice-drills me on my own CV claims with faster-whisper local transcription, OpenAI gpt-5.4-tiered pipelines, in-app Marimo-style labs with an LLM grader, SM-2 spaced repetition, daily 8 AM IST Gmail digest. ~$0.04–0.09 per drill.
End-to-end job search automation: Gmail alert ingest, 17-portal API scan, 9-check zero-token prefilter, Haiku 4.5 smart filter, Reddit + HN "who is hiring" radar, Chrome extension JD capture, auto-resume tailoring, dashboard on :8001. Hourly systemd timers, Haiku ~$1–3/mo, Apify $0 after a free-direct backend rewrite.
Full-stack career intelligence platform — FastAPI + Next.js + SQLite. Resume parsing (PyMuPDF + LLM), LinkedIn scraping, GitHub analysis, LLM-powered skill normalization, verification quizzes, gap analysis, job match scoring.
Tinder-style photo culling tool — swipe through ~6,000 wedding photos streamed straight off Google Drive (no re-uploads, no double-storage). Keep / discard / favorite, persistent verdict log, resumable session. Built so my family could finally finish picking.
A NeRF-based approach to occlusion and illumination challenges in AR scene augmentation. Extended individual NeRFs as composable blocks via a Generalizing module for intermediate merged renderings. Poster at CSE Research Symposium 2023.
Ubuntu Server 24.04 self-hosting 10+ Dockerized services: Jellyfin, qBittorrent, Sonarr, Radarr, PostgreSQL, Redis. UFW firewall, Tailscale VPN mesh, Samba shares. Gaming PC with RTX 4060 Ti 16 GB running Ollama for local LLM inference.
I like building the unglamorous middle of AI systems — the retrieval that doesn't hallucinate, the eval that tells the truth, the integration that survives a real customer. Models are the easy part.
My stack is breadth over depth on purpose: a production LLM bug can live in the prompt, the retriever, the API surface, the queue, the DB — and I'd rather be the person who can walk the whole trace than one who owns just one layer.
Nights and weekends I run fine-tunes on a 16 GB card in my bedroom and write down the regressions. I'd rather ship a boring mechanism that works than a clever one that needs an explanation.
If you're working on AI systems that need to actually ship — RAG, agents, fine-tuning, real-time voice AI — I'd love to talk.