Building production-grade AI systems — multi-agent LLM orchestration, multimodal RAG, speech & video ML, legal AI, and civic technology. 3+ years of hands-on ML engineering, plus open-source and applied research.
Flagship production systems, research implementations, and agentic AI architectures.
Isochrony-aware video dubbing pipeline translating English to 11 Indic languages with zero-shot voice cloning. The v3 fine-tune constrains translation to a strict phoneme budget, cutting length error 79% (0.495→0.103 rel.) while improving translation quality (chrF++ 22.0→30.1).
A multi-account email agent consolidating Gmail, Zoho, Yahoo & Outlook into one AI-enriched inbox — a single MCP server that is itself an MCP client, fanning out to each provider and re-exposing 17 unified tools with self-healing child process reconnection. 49 passing tests.
Production civic-issue reporting platform: residents map potholes, garbage and drainage across Noida and NCR, with server-rendered social share cards, a two-tier trust model, and a hardened upload pipeline — all running for about $3/month.
An open-source bounty & issue aggregator whose moat is telling the truth about a bounty — including when the truth is "don't bother." A pure-function, three-layer trust engine rejects engagement farms and crypto-deposit scams before any network call; live-validated when a real farm's admitted count dropped from 20 to 0.
Ten cooperating AI agents that turn long podcasts into scheduled short-form videos — a provider-agnostic LLM layer with automatic fallback, a learning feedback loop, and a QA gate. Runs unattended on free providers. MIT-licensed.
Paste a YouTube channel URL and it reverse-engineers the style, writes an original script in that voice, narrates and illustrates it, and compiles a finished MP4 — for $0. A three-tier LLM failover, self-healing JSON validators, and a zero-network procedural generator guarantee output even when every API is down.
AI-powered Indian Knowledge System oracle enabling semantic querying of Vedic texts, Upanishads, and ancient Indian manuscripts using multilingual RAG & graph-based retrieval.
Interactive dashboard for valuing football actions using the VAEP framework. Analyzes match events across Top 5 European Leagues and World Cup 2018 to rank players and visualize goal-scoring sequences.
Interactive Expected Goals (xG) simulator with pitch-mapping and heatmap visualization. Evaluates scoring probability in real-time.
Advanced semantic matching engine using Sentence-Transformers. Features skill extraction and gap analysis for job applications.
CNN-based mobile diagnostic tool for identifying crop diseases in real-time. Achieved 98% accuracy on the PlantVillage dataset.
Real-time weather tracking with a secure proxy layer. Implements global rate limiting and intelligent caching to manage API quotas.
Explore and test live production ML models deployed on Hugging Face Spaces. Profile: huggingface.co/nirmitk
AI-powered Indian Knowledge System enabling semantic search & graph-based exploration over Vedic texts and ancient Indian manuscripts with multilingual embeddings.
Action-valuation dashboard quantifying offensive and defensive match events across Top 5 European Leagues and World Cup 2018 using the VAEP framework.
Interactive pitch coordinate mapper and statistical scoring probability estimator powered by a Random Forest Classifier trained on European league shot events.
Deep semantic matching and skill gap analyzer comparing candidate resumes against job descriptions using Sentence-Transformers (SBERT all-MiniLM-L6-v2).
Deep CNN vision diagnostics identifying 38 crop foliar diseases in real-time from leaf imagery with 98% validation accuracy on the PlantVillage dataset.
View the complete catalog of open-source datasets, model weights, and live web applications hosted under the @nirmitk Hugging Face organization.
Deep-dives on the engineering and research behind my projects — written to be read, not skimmed. From production systems to primary-source history.
AI dubbing drifts because translations aren't the same length as the original. The industry fixes it in the audio; I fixed it in the translator — cutting length error from 49% to 10% on a fine-tuned Llama-3.1-8B, with the full training telemetry left in.
v2.5 measures length in real phonemes and gates every line on meaning, so it passed 42 of 44 test segments across 11 languages. This is how it runs today on a free Kaggle T4, and what the move to an async GPU API has measured so far, with projections labelled as projections.
A GPU I didn't pay for, running an app I could open from my phone, that I could kill from anywhere — a tunnel out, a public message channel back, and a kill switch. Plus the meta-tensor bug that cost the most time.
Detection asks what just happened; anticipation asks what's about to. My SoccerNet 2026 scores say the model knows what is coming far better than when — and the decomposed metric shows exactly which half broke.
The architecture: one MCP server that is also an MCP client, consolidating four mail providers behind a single 17-tool surface — with exactly one custom component.
An MCP connection is a child process, and child processes die. The self-healing reconnection — isolation, drop-detection, and a cooldown — that makes it survivable, with two production bugs pinned in tests.
Empty JSON from an 80-token task with a 512 budget. On a thinking model the fix for a truncated answer is a bigger budget — and it fails completely silently. Floor it, double on retry.
I verified four real Algora bounties live and found every winnable-looking one saturated or already paid. The honest move was to report the negative result, not manufacture a win — which is exactly what the product promises its users.
Ideation and research behind a trust-first bounty aggregator: the polluted label space, the bounty-plaza crypto-deposit scam we reverse-engineered, and the four falsifiable hypotheses that shaped the build.
The async FastAPI + Next.js architecture, a natural-key data model, and a three-layer trust engine that rejects farms and scams before any network call — validated by unit tests and a live re-probe that dropped a farm's admitted count from 20 to 0.
The end-to-end bounty lifecycle run live: public boards on hot orgs are saturated red oceans. Why the failure created GitScout's second product axis — a viability signal — and why no completed claim was manufactured.
How ten specialised AI agents hand off work like a production company — a provider-agnostic brain with automatic fallback, prompt-injected craft playbooks, a learning loop, and a QA gate that blocks broken output.
One channel URL in, a finished narrated video out — for $0. The engineering that matters isn't any single model but the resilience layer: a three-tier failover cascade, self-healing validators, and a procedural generator that guarantees output when every API fails.
A history-to-policy argument from 18th-century primary sources: India's mechanization bottleneck has always been access to divisible draft power — not tools, cost, or awareness — and it still is.
Making a JavaScript map describe itself to WhatsApp crawlers, trusting strangers just enough to let them post, and hardening an upload path that assumes every phone photo carries home GPS — for ~$3/month.
The production v2 system — how it constrains Gemini translation to a phoneme budget, preserves the soundstage with source separation, and clones voices at a controlled duration, all on a free Kaggle T4.
I'm a certified AI/ML Engineer with a M.Tech in Modeling & Simulation and 3+ years of industry experience building intelligent systems across NLP, LLMs, multimodal ML, and agentic AI.
My work spans production deployments at Chetu India and BLU PARROT — including an agentic event-commerce chatbot, a Video-to-SOP multimodal pipeline, legal AI for Indian law, sports action anticipation research, and a participant in the SoccerNet 2026 Challenge.
Most recently AI/ML Software Engineer at Chetu India (through mid-2026). Now available for freelance, consulting, and research collaboration — remote — and open to full-time roles from September 2026.
From classical ML to cutting-edge multimodal AI — built and shipped in production.
LangGraph multi-node agents, RAG variants (Naive/LightRAG/CRAG/VideoRAG), LangSmith tracing, OpenAI & Gemini APIs, FastAPI.
CNNs, Transformers, VideoRAG, ImageBind multimodal embeddings, OpenCV, FUTR anticipation models, T-DEED temporal spotting.
BERT variants (InLegalBERT, XLM-RoBERTa, IndicBERT), FasterWhisper ASR, multilingual NMT, text classification, summarization.
FAISS, HNSWLib, Neo4j, NetworkX, HuggingFace, W&B, Python scraping, Azure OCR, data engineering pipelines.
Expected Goals (xG) modelling (89.77% accuracy), FAANTRA ball action anticipation, SoccerNet 2026 Challenge participant, pass network analysis.
LegalGPT: case summarization, 150+ contract types classification, compliance AI, Opennyai, InCaseLawBERT fine-tuning.
Building AI that ships — from startup chaos to enterprise scale.