Kawchar Husain@kawchar85
2
Events attended
2
Submissions made
Bangladesh
3+ years of experience
About me
ML Engineer focused on GenAI, LLMs, multimodal systems, computer vision, and AI agents. I enjoy turning research ideas into working products—from model fine-tuning and inference pipelines to agentic systems, APIs, and end-to-end AI applications. I work primarily with Python, PyTorch, FastAPI, and modern LLM tooling, with experience in fine-tuning, RAG, multimodal integration, and production-oriented ML systems. Beyond ML, I’m a Codeforces Expert and an ICPC Asia West Continent Finalist, with experience in competitive programming, problem setting, research, and teaching. I especially enjoy hackathons and projects where there is room to experiment, build quickly, and solve problems in unconventional ways.
🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
Roadmap Canary
Roadmap Canary addresses a blind spot in traditional CI: tests can prove that today’s software still works, but they cannot tell you when a proposed change quietly makes an already committed future capability much harder to build. A codebase can stay completely green while a refactor introduces vendor-specific coupling, removes an abstraction, or turns a previously small future change into a major architectural rewrite. Our solution treats selected future roadmap commitments as executable engineering artifacts. For each architecture-sensitive capability, the team approves a Future Contract that defines what must remain demonstrably possible, the verification commands and trusted inputs that count as evidence, and the proof budget that still qualifies as a small viable path. Roadmap Canary then captures a verified witness: a small, disposable executable proof that the capability is achievable from the current architecture without becoming production code. When a branch or pull request is analyzed, Roadmap Canary replays the witness on both the approved BASE and the proposed Change. If the witness still works, the capability remains SAFE. If it already fails on BASE, the Canary is STALE and Roadmap Canary does not blame the Change. If it passes on BASE but fails on the Change, Roadmap Canary invokes IBM Bob as a bounded Rescue engineer. Bob receives the human-approved Future Contract, repository context, and witness failure evidence, then searches for a replacement executable path inside an isolated Git worktree. Bob only proposes code. Roadmap Canary independently decides whether that proposal counts as valid evidence by running deterministic verification: existing tests, type checking, future-capability tests, protected verifier inputs anchored to BASE, and proof-budget checks such as changed files, added lines, and new dependencies. This produces clear engineering outcomes: SAFE when the known path still works; PATH CHANGED - SAFE when the old path breaks but Bob finds a new verified path within budget; ROADMAP RISK when no replacement proof is accepted within the bounded Rescue attempt; and STALE when the stored Canary can no longer be attributed to the proposed Change. The primary users are platform and architecture teams, staff and principal engineers, and engineering organizations with committed roadmaps and shared codebases. They interact through a focused workflow: choose a repository and future capability, select BASE and Change, run analysis, then inspect the replay evidence, Bob Rescue candidate, deterministic checks, proof budget, and final verdict. What makes Roadmap Canary different is that it preserves capability, not one speculative implementation. It turns one-off architectural spikes into persistent regression sentinels and moves the discovery of future architectural cost from months later, when a feature finally starts, to the change that introduced that cost.
27 Sep 2026
.png&w=828&q=75)
AIchemist-agent
AIchemist-agent is a general-purpose AI agent that takes a batch of open-ended natural-language tasks and returns a complete, judge-ready answer for every one of them. It handles the full spread of general-purpose work without any task-specific configuration: factual questions, multi-step arithmetic and quantitative reasoning, constraint and logic puzzles, sentiment analysis, named-entity extraction, summarisation under strict length and format demands, code generation, and debugging of broken code. Nothing is hardcoded or pre-baked for known inputs — the agent reads whatever tasks it is given at runtime and answers them fresh, in English, on unseen variants. Reliability is treated as a first-class requirement rather than an afterthought. The agent honours the literal output contract a prompt asks for: exactly three bullet points means three, "under 130 characters" means under 130, "a JSON array" means a raw parseable array with no prose or code fences around it. Answers are validated before they are emitted, and an answer that does not satisfy the request is corrected rather than shipped. Long, complex tasks never degrade into a truncated fragment or a stream of visible deliberation. It is also built to survive the operating environment it is graded in. Every task ID receives an answer even if an individual task fails, the output file is always valid JSON, and the process exits cleanly. All credentials and model identifiers are supplied by the environment at runtime — nothing is baked into the image, and no specific model is assumed to exist. The whole run is bounded by an internal time budget that keeps it comfortably inside the harness's limits, on a 4 GB, 2-vCPU machine. The result is an agent that is broadly capable, strictly compliant, and markedly cheaper to run than a naive one — on a 110-task internal benchmark it answered every task while spending a small fraction of the tokens a direct-to-frontier-model baseline consumed.
13 Jul 2026
