Fine-tuning a Mistral-7B classifier (QLoRA) to turn unstructured infrastructure log telemetry into structured root-cause analysis. Scaling a hand-curated dataset to 500 examples with a 10% holdout, evaluated against a dedicated harness.

→ Phase_2/LoRA

Every project below is raw Python. No LangChain, no LlamaIndex, no third-party agent framework, no API integrations.

ReAct Agent A ReAct (Reason → Act → Observe) loop running a local Llama-3.2-1B-Instruct GGUF model via llama-cpp-python, CPU-only. Hand-rolled action parser, tool dispatcher, and a circuit breaker that halts on any repeated (action, param) pair anywhere in run history — not just the immediately preceding step.

KV Cache VRAM Calculator Derived the KV cache memory formula from first principles, then verified it empirically against Llama-3.2-1B-Instruct on a Colab T4 — including the Grouped Query Attention trap: using num_attention_heads instead of num_key_value_heads overestimates cache size by exactly the GQA ratio (4x for this model).

Micrograd Karpathy-style scalar autograd engine, implemented line by line: a Value class with __add__, __mul__, tanh, and a manual backward pass verified against finite-difference gradient checks.

OpenHands/software-agent-sdk #4435 — merged Fixed a bug where the condenser's max_tokens wasn't inheriting the LLM's effective max input tokens, silently truncating context on specific launch paths. Merged by a core maintainer.

Languages

ML / Fine-Tuning

Local Inference

Tools