Research infrastructure, developer tools, and engineering projects. Each record separates scope, supported evidence fields, date, methods, and the full case record.
ScopeMeasure which exact session facts survive coding-agent context compaction instead of treating memory loss as an opaque summarization problem.
Role
Designed and implemented the typed event schema, compaction policies, instrumentation, and bounded experiment harness in an OpenCode fork
Outcome
On the first bounded GPT-5.5 noisy-compaction fixture, the precision-frontier policy kept full target recall with 852 compaction-input tokens; solve-rate evidence is not yet established.
ScopeEvaluate coding-agent repairs with bounded context, sandbox validation, and retained evidence instead of accepting a plausible patch at face value.
Role
Designer and implementer of the repair harness, task adapters, validation flow, artifact schema, and readiness gates
Outcome
The current DeepAgents public-issue smoke lane validates three of three focused pytest/Requests tasks while retaining traces, validation output, and failure reports.
ScopeHelp coding agents identify the repository evidence they need before editing, then measure whether that guidance improves real target-file reading.
Role
Creator and maintainer of ctxhelm; designer of the companion HelmBench evaluation harness
Outcome
Released ctxhelm v2.4.0 through GitHub archives and Homebrew; a completed 10-task RefactoringMiner Codex matrix preserved solve and validation rates while improving retrieval-oriented measures.
ScopeProcess one billion temperature records correctly on Apple Silicon while avoiding the mmap fault path that dominated full-size runs.
Role
Sole implementer and benchmark author
Outcome
Passed 12 official fixtures and matched generated 100M, high-cardinality 10M, and full 1B outputs; the promoted full-size path won all five paired warm wall-time runs on the stated Apple Silicon machine.
ScopeUnderstand tensors, automatic differentiation, neural-network layers, and training loops by rebuilding their core mechanics in a small Python framework.
Role
Independent learning build
Outcome
The milestone sequence and intended interfaces are documented; implementation evidence has not yet been published.
ScopeRender 3D scenes from first principles while making ray intersections, lighting, cameras, materials, and scene parsing inspectable in C++.
Role
Implemented and documented the C++ ray-tracing course project
Outcome
The repository and case record cover ray-object intersection, lighting, camera control, scene description, and extension paths for acceleration and advanced materials.
Status
Completed BUET computer-graphics course project
Documented
MethodsC++ · Computer Graphics · Physics Simulation · 3D Rendering · Ray Tracing
ScopeImprove image-caption generation by combining visual attention with contextual language representations and measuring the change against a baseline.
Role
Co-developed the model extension, experiment pipeline, and technical analysis
Outcome
Reported 36% CIDEr and 43% BLEU-4 improvements over the project baseline.
ScopeReduce ticket fraud and abusive resale while preserving verifiable ownership, controlled transfer, and venue entry.
Role
Co-developed the architecture and contract logic, focusing on lifecycle rules, transfer constraints, and validation flows
Outcome
The team was selected as a Blockchain Olympiad Bangladesh 2021 finalist with a prototype covering ticket issuance, marketplace resale, and validation flows.
