Lead Autonomous Systems & Punjabi NLP Architect
Building the future of local, private, and self-evolving multi-agent intelligence.
I am a software engineer and AI researcher specializing in autonomous agent architecture, self-evolving codebases, and Gurmukhi NLP. My work focus is on pushing the boundaries of private local execution on Apple Silicon (M-series GPUs), implementing multi-agent swarms that research, plan, code, and self-debug in real time.
An autonomous, local multi-agent system comprising 22 specialist agents synced via an asynchronous event bus and DAG task scheduler.
- Self-Evolution Pipeline: Auto-scans python files, runs its test suite (109 tests passing), refactors modules, and applies updates in a secure sandbox.
- Multimodal Capabilities: Integrates Whisper-based voice (STT/TTS) and vision agents.
- Punjabi NLP & guardrails: Gurmukhi-aware prompts and ethical guardrails (Dharam, Sach, Daya).
An autonomous scientific discovery and medical research engine designed to synthesize hypotheses, simulate scenarios, parse arXiv & PubMed databases, and draft comprehensive papers.
☬ Sovereign Foundation Models (Nam-toon Studio)
Punjabi-first models trained from scratch on a laptop (Apple Silicon). Numbers below are the ones measured on each model card — small models are fluent in Punjabi but still weak on facts.
- 🙏 Sevak-97M Sovereign: 97.2M-parameter Punjabi + English LM (16K Gurmukhi BPE). 14.7% on a decontaminated 143-question Punjabi factual test — honest about its hallucinations.
- 🌱 Sevak-24M Sovereign v2: smallest Sevak (29.0M parameters released), same recipe; 8.4% on the same test.
- 🧠 Sahaj-86M: 86M hybrid selective-SSM + attention Punjabi LM with a retrieval (RAG) layer and a confidence/abstention probe — 12% factual QA unaided, ~40% with retrieval.
- ⚡ Sahaj-30M: ~29.6M byte-level experimental Punjabi LM (no tokenizer). Early research model, not yet benchmarked.
- 📚 Gurmukhi datasets: grammar correction, Gurbani / Mahan Kosh study corpus, Punjabi STEM chain-of-thought, synthetic Punjabi TTS speech — see each card for how the data was made.
Sevak-Hybrid (HF): a router of 1.5–2B open Qwen specialists plus adaptive test-time compute that matches Microsoft's Phi-4-mini (3.8B) on held-out GSM8K (93.5% vs 92.5%), MMLU (71.3% vs 70.7%) and HumanEval (59% vs 56%), every system run in the same harness on one laptop. A system, not a new model — no weights trained; gaps are within noise.
| Layer | Technologies & Tools |
|---|---|
| Languages | Python, JavaScript, Swift, Bash, Gurmukhi NLP |
| Agentic Frameworks | Swarm, CrewAI, Custom EventBus, DAG Task Scheduler |
| Local LLMs | Ollama, AirLLM (32B Escalation), local LoRA adapters |
| Vector & Databases | FAISS, SQLite, SQLite-VSS |
| Multimodal STT/TTS | Whisper, Custom TTS caching |
| Deployment & Ops | Docker, Makefile, GitHub Actions, Apple Silicon (Metal GPU) |
Connect & Collaborate:
GitHub · Hugging Face (Nam-toon Studio) · LinkedIn · gurpreetsingh5523@gmail.com
☬ ਵਾਹਿਗੁਰੂ ਜੀ ਕਾ ਖ਼ਾਲਸਾ ॥ ਵਾਹਿਗੁਰੂ ਜੀ ਕੀ ਫ਼ਤਹਿ ॥