SaicharanRamineni
Software Engineer specializing in
Agentic Systems, Distributed Architecture, and High-Performance Infrastructure..
Technical
Highlights
A selection of systems developed at the intersection of machine intelligence and high-performance engineering.
Sai
Voice-native macOS co-pilot that drives the operating system through pure vision — no DOM, no accessibility tree. A hybrid Nova Lite/Pro routing pipeline classifies intent in sub-second latency, while an Annotated Vision Pipeline with LANCZOS resampling maps logical agent coordinates directly to physical UI pixels.
Project Sentinel
Enterprise-grade real-time Voice Anti-Fraud Guardian. Leverages ultra-low latency Llama 3.3-70b inference on Cerebras and full-duplex WebSockets to intercept live audio streams, analyze conversational semantics, and deploy immediate active-defense voice warnings.
Crisis Averted
A 3D geospatial intelligence platform that maps humanitarian funding mismatches in real time. Features Pablo, a sovereign AI voice orchestrator that allows users to navigate a 3D globe and analyze data using natural language.
AeroForge
Autonomous Generative Engineering Platform driven by a Gemini 3 Pro multi-agent swarm. Orchestrates real-time physics simulations across distributed Vultr NVIDIA H100 clusters via Apache Kafka to design, test, and optimize production-ready mechanical systems with zero human intervention.
Vectorized 2048
Massively parallel Deep Reinforcement Learning environment built entirely in JAX. Engineered a custom Dueling Double DQN and a branchless, differentiable physics engine to simulate 4,096 simultaneous games natively on GPU/TPU hardware.
Complete Archive
Explore the full breadth of work spanning ML research, systems engineering, and interactive visualizations.
Infrastructure
& Logic
Expertise spanning the full stack of high-performance computing, from low-level systems to advanced neural architectures.
AI, ML & Accelerated Compute
Data Engineering & Cloud Infrastructure
Real-Time Web & Application Development
Open
Source
Accepted upstream contributions to the libraries I build on — diagnosing real bugs and shipping fixes maintainers review and adopt.
Prevent process crashes across nine sparse optimizer kernels
A lower-rank gradient could make TensorFlow's sparse optimizer kernels access a nonexistent tensor dimension and abort the process. Added explicit rank validation across all nine affected CPU kernels plus regression coverage; maintainers approved the patch, and it landed during TensorFlow's repository-sync recovery.
Fix float32 tf.math.erfinv precision loss near ±1 in the eager path
tf.math.erfinv in eager float32 drifted up to ~1% from the true value near ±1 while the XLA path stayed accurate to ~7 digits: the eager functor's ndtri(0.5*x + 0.5) * sqrt(0.5) snaps the argument onto the coarse float grid around 1.0, destroying the tail distance ndtri needs. Added a float specialization (scalar + vectorized packetOp) evaluating erfinv directly via the Giles approximation with w = -log1p(-x*x) — matching xla::ErfInv32, preserving the ±inf/NaN edge cases and leaving the accurate double path untouched — with a regression test against the float64 ground truth.
Fix a segfault in TensorListSetItem on out-of-range indices
tf.raw_ops.TensorListSetItem segfaulted on an out-of-range index: a huge value passed as int32 wrapped to -1, which promoted to a giant size_t in the bounds check, took the resize branch, then indexed the tensor list out of bounds. Validated the index — rejecting negatives and indices past max_num_elements — and cast before the +1 to avoid overflow, with a regression test.
Make Studio setup work from paths containing spaces
uv 0.11 truncates -c/-r file paths at the first space, so `unsloth studio` setup failed when the repo lived under a path like 'Open Source/'. Extended the _uv_safe_path helper to hand uv a space-free temp copy on macOS/Linux — with atexit cleanup, a safe fallback, and tests.
Fix UTF-8 crashes in Studio's code-execution tool on Windows
Studio's Python code-exec tool crashed on Windows whenever model output held non-ASCII (arrows, CJK, emoji), because the script write and pipe read fell back to cp1252. Forced UTF-8 across the subprocess I/O and added a round-trip regression test.
Shim the module vLLM deleted so fast_inference survives upgrades
Added a sys.meta_path finder that supplies a no-op stub for the module vLLM dropped in v0.22 — only when the real one is gone — so users who upgrade unsloth ahead of unsloth-zoo don't crash. Defensive companion to the root-cause fix.
Surface config-file guardrails in the usage dashboard
Config-defined guardrails live in memory, not the DB, so the proxy's usage endpoints mislabeled them as 'Custom' and 404'd on detail lookups. Resolved them by generated ID or name and surfaced the real provider (e.g. presidio), with regression tests.
Force store=false for OpenRouter Responses API
OpenRouter rejects Responses API calls unless store is false-or-absent, but LiteLLM forwarded store: true/null from clients. Forced it to false before the OpenAI-compatible transform, covering both the explicit and omitted paths with tests.
Enforce waitable-set traps for sync stream/future cancel and subtask.cancel
Wasmtime enforced the component-model waitable-set trap for synchronous read/write but missed sync cancel-read/write and subtask.cancel, letting spec-invalid cancels complete. Hoisted the guard into a shared helper, wired it through the sync cancel paths, and added a wast regression covering future/stream cancel-read/write plus subtask.cancel.
Fix broken f-string in dataloader resume warning
A checkpoint-resume warning concatenated two string literals but only f-prefixed the first, so the dp rank id it exists to surface printed literally as {self._rank_id}. Added the missing prefix plus a regression test covering the missing-rank branch.
Fix async cancel ordering to leave the waitable set first
A synchronous stream.cancel-read ran while the inter-task wakeup stream was still in the task's waitable set — which the canonical ABI requires to trap, faulting on runtimes that enforce it. Reordered removal before the cancel to match the established unregister-then-cancel pattern.
Fix fast_inference crash on vLLM ≥ 0.22
Root-caused a ModuleNotFoundError that broke GRPO + fast inference on Colab after vLLM removed its tokenizer group, and guarded the unsafe import so older and newer vLLM both work.