Hey, I'm Sahil Kumar, an AI/ML engineer in India. I build AI systems from scratch: speech pipelines, RAG, TinyML, and autonomous robots.
A pipeline that compiles multi-speaker Hindi/English lecture and podcast audio once into a frozen canonical transcript where every word carries start/end timestamps and a speaker label, then serves subtitles, cited chapter summaries and moment-level search from that single artifact. faster-whisper large-v3 with an OOM fallback ladder, a local WeSpeaker ONNX plus spectral-clustering diarizer (9.1% DER), and hybrid BM25 + BGE-M3 retrieval with cross-encoder rerank (80% Recall@5 on 60 gold questions), all sequenced to fit a 4 GB laptop GPU at 2.1 GB peak VRAM. Evaluation is fully generated and reports the misses honestly, including alignment drift and Hinglish WER.
A Go control plane (ffmpeg HLS segmenting, Redis work queue with worker leases, idempotent retries) dispatching to Python GPU workers, turning an hour of multi-speaker Hinglish cricket audio into structured, citable content in ~7 minutes on one consumer GPU. faster-whisper large-v3 plus WhisperX alignment and pyannote diarization; a lexicon bias pass cut entity WER from 31% to 12%. Hybrid BM25 + dense retrieval over Qdrant with cross-encoder rerank, every answer citing a word-level timestamp and abstaining below the evidence threshold. Shipped as FastAPI microservices on Kubernetes via GitLab CI, gated by a 40-clip golden set that fails the build on a >3% regression.
A personalized nutrition engine that classifies food with EfficientNet-B0 and reports calories as uncertainty bands, not false-precision numbers. Stores biometric time-series in SQLite, pulls activity via Google Fit OAuth, and surfaces only statistically honest correlations (Spearman with Benjamini-Hochberg FDR control). FastAPI service that picks honest over impressive.
A multimodal Retrieval-Augmented Generation system built in pure Python with no framework glue and no API keys. FAISS vector search, CLIP image retrieval, and local Ollama LLMs behind a FastAPI service, with every stage (chunking, embedding, retrieval, generation) readable top to bottom.
A misinformation and AI-content detector that fuses four independent signals (a DeBERTa-v3 text classifier, C2PA/EXIF metadata, SynthID, and image forensics) into one calibrated confidence with an uncertainty band, instead of a binary real-or-fake verdict. FastAPI plus ONNX, deployable to AWS Lambda.
A mobile robot running three concurrent real-time loops (vision, voice, hardware control) in Python, with a fully local voice pipeline: openWakeWord hotword, Faster-Whisper STT on CUDA, Gemma via Ollama, and Piper TTS. ESP32 firmware on FreeRTOS drives an L298N H-bridge over TCP, with a Flutter app for live telemetry and manual control.
B.Tech, Electronics & Communication Engineering (AI/ML Specialization)
Selected for the international final round in Indonesia with an AI-based cloud anomaly detection project.
Built a full-stack e-commerce site for a jewellery business and reached the finalist round.
Auto-triaging incidents with SigNoz webhooks and an LLM, and the wall almost nobody gets past.
How do you run a resource-heavy middleware like ROS2 on a tiny microcontroller like an ESP32? Enter micro-ROS and XRCE-DDS: the client-agent bridge that offloads heavy discovery to the host.
Models get bigger, context windows grow, memory improves. But the underlying mechanics haven't changed. Understanding them is more useful than waiting for the next upgrade.
I just watched a talk by Andrej Karpathy that fundamentally shifted how I see the future of coding. He argues that we're in the middle of the biggest change in 70 years. He calls it Software 3.0.
Whether you're working on the next generation of robotics, advancing AI architectures, or just want to talk Linux/open-source — my inbox is always open.