I build LLM agents and the evals that show whether they work. B.Tech AI & Data Science @ Shiv Nadar University · 2027
Open Source — OpenHands/software-agent-sdk
Harness-level bug fixes in the OpenHands agent SDK, each shipped with regression tests.
- #4435 · merged — The context condenser's
max_tokenswasn't inheriting the LLM's effective max input tokens, silently truncating context on specific launch paths. - #4566 · merged — The global
load_memorypreference reached only one of three conversation launch paths. Centralized the fix; 5 regression tests. - #4876 · merged —
load_skills_from_agent_server()discarded the caller'sAgentContext. A follow-up bug I found and filed after #4566; 5 regression tests. - #5690 · in review — Concurrent plugin requests could corrupt
.installed.json. File lock plus atomic write.
RCA LoRA Scaling Study — does more fine-tuning data help? Controlled QLoRA ablation on Mistral-7B: training set scaled 20 → 474 examples, with LoRA config, quantization, and prompt format held fixed. Scored by a 3-tier harness I built: JSON validity, embedding similarity, calibrated LLM-as-judge. The first two tiers called it a near-tie. The judge scored the 474-example model lower (2.00 vs 2.21, n = 53 held-out), and manual cross-checks confirmed fabricated technical content in 8 of 9 cases. More data didn't help at fixed adapter capacity.
ReAct Agent — a raw-Python ReAct loop on a local Llama-3.2-1B-Instruct GGUF via llama-cpp-python, CPU-only. No agent framework: hand-rolled action parser, tool dispatcher, step budget, and a circuit breaker that halts on the first repeated (action, param) pair anywhere in run history.
KV Cache VRAM Calculator — the KV cache formula derived from first principles and verified on a T4 against Llama-3.2-1B-Instruct. It matched direct cache-tensor inspection exactly (25.31 MB), while torch.cuda.max_memory_allocated() reported 48.10 MB. Covers the GQA trap: using num_attention_heads instead of num_key_value_heads overestimates the cache 4×.
Siamese Bi-Encoder — a toy Sentence-BERT setup trained from an empty embedding table. Perfect test AUC across 3 seeds hid a test MSE 750–1,500× the train MSE.


