K Ashraf Ahmed

weightedRoundRobin Load-Balancing Algorithm

zalando/skipper·PR #4124·+599/−4·3.1K+ stars

Designed and shipped a smooth weighted round-robin algorithm with dynamic health-derived endpoint weights. Defended the locking design under maintainer review with a self-built C benchmark harness sustaining ~31k req/s at 1,000 endpoints — the harness was kept in the repo for testing all algorithms.

lifoWithBody / lifoGroupWithBody — Concurrency-Limit Fix

zalando/skipper·PR #4188·+275/−17·3.1K+ stars

lifo released its concurrency queue slot as soon as the backend sent response headers, so a backend streaming a long body left the slot free and the route admitted past MaxConcurrency. Added lifoWithBody and lifoGroupWithBody that hold the slot until the body finishes streaming, using a separate state-bag key so a plain lifo() on the same route can't pop the release function early. The regression test fails against the old filters — they wrongly return 200 where the new ones correctly return 502 on queue timeout.

mobilize — Wave Dispatch Agent + MCP Server

CALLE-AI/awesome-phone-call-agents·PR #98·+6,714·CALL-E stars

Shipped an agent app and MCP server that treats a model's stated confirmation as a noisy signal rather than a fact — scoring call evidence against historical show-rate before it counts. A 300-trial evaluation harness against a synthetic population with known ground truth measured 94.6% accuracy for the calibrated policy vs 87.7% for trusting every stated yes. Crash-safe dispatch via a write-ahead ledger, verified with a real SIGKILL'd subprocess. 89 tests.

Prometheus Native Histogram Support

zalando/skipper·PR #4108·+245/−146·3.1K+ stars

Added opt-in Prometheus native histograms behind a flag, with bucket limits set per the Prometheus and OTEL recommendations to bound memory. Classic histograms are still emitted alongside them, so existing dashboards and alerts keep working. Fixes #3898.

Per-Data-Client Load Latency Metrics

zalando/skipper·PR #4087·+76/−1·3.1K+ stars

Added instrumentation to measure route-loading latency per data source, giving SRE teams visibility into which integrations are slow.

TLS Certificate Matching Documentation

etcd-io/website·PR #1171·+6/−3·48K+ stars

Clarified security documentation for certificate-matching semantics across v3.5–v3.7, resolving a 14-month-old confusion issue about wildcard support. Merged by a core maintainer.

Go Toolchain Security Updates

etcd-io/gofail·PR #149·+2/−2·CNCF stars

Patched three CVEs by bumping Go to 1.25.11 as part of etcd organization's tracked security effort.

01

Kiln

LLM Inference Engine

An LLM inference engine written in C++ with custom CUDA kernels for attention, RMSNorm and RoPE. It serves SmolLM2-135M at 219–231 tokens/s on a GTX 1660 Ti through an OpenAI-compatible API (one request at a time), with final logits within ~6×10⁻⁵ of Hugging Face on two checkpoints. In a CPU benchmark, KV caching sped up generation 11.2× over full recomputation, and a paged KV cache shares common prompt prefixes copy-on-write. In a GPU microbenchmark, INT8 matrix multiplication runs 1.60× faster than FP32. 136 tests.

• C++
• CUDA
• Python
• PyTorch
• LLM Serving
• Quantization

02

Consensa

Distributed Vector Database

A distributed vector database in Go: hand-written Raft consensus, an MVCC storage engine, HNSW vector search, range sharding with live splits, and cross-shard transactions across two 3-node replica groups. A chaos harness kills processes and partitions the network, then checks the histories for linearizability. Quorum intersection and range-key ownership are model-checked in TLA+. Local HNSW search returns 98.6% of the true top-10 matches in 156 µs on 5,000 synthetic vectors.

• Go
• Raft
• HNSW
• gRPC
• TLA+
• Distributed Systems

03

Anamnesis

Memory for AI Agents

A memory layer for AI agents on CockroachDB, built for the CockroachDB × AWS hackathon and deployed on AWS Lambda. Beliefs carry validity intervals, so the agent can answer what it believed last week. New beliefs are checked against old ones and contradictions supersede instead of overwriting, and every write is audited in the same SERIALIZABLE transaction.

• Python
• FastAPI
• CockroachDB
• AWS Lambda
• LLM Agents

04

Aegis

Predictive Observability Platform

Failure prediction and root-cause analysis for microservices. A Kafka event-time correlator traced the root cause in all 78 simulated incidents, and a gradient-boosted model warned a median 193 seconds before failure on 18 held-out simulated episodes, with shadow scoring and PSI drift monitoring around it.

• Python
• FastAPI
• Kafka
• Scikit-Learn
• Prometheus

05

ClearText API

Async ML Inference Platform

A FastAPI service that takes BERT inference off the request path: the API returns a task ID while batched Celery workers run the model. Repeat inputs are served from a model-versioned Redis cache, requests get a 503 with Retry-After when workers are down or the queue is full, and failed tasks land in a replayable dead-letter queue.

• FastAPI
• Celery
• Redis
• PostgreSQL
• BERT

06

Go + Java Auth Platform

Go · Java · Spring Boot

A Go auth service with rotating refresh tokens, reuse detection and a Redis JTI blacklist, consumed by a Java/Spring Boot API that validates each request over gRPC behind a token-hash Caffeine cache and a fail-closed Resilience4j circuit breaker. 61 tests (Testcontainers and in-process gRPC); k6 load test at ~92 req/s with 6 ms p50 across both services.

• Go
• Java
• Spring Boot
• gRPC
• JWT
• Redis

07

Radar–Camera Fusion for Real-Time Drone Detection

Python · PyTorch · YOLOv8n

Confidence-weighted late fusion of YOLOv8n detections with radar-inspired motion features, lifting precision from 0.91 to 0.94 (F1 0.92) at 8.3 FPS on CPU, with no GPU needed at inference time. An LSTM trajectory forecaster cut RMSE from 1.78 to 1.64 against a GRU baseline. Basis for two IEEE papers (ASIACONF 2026, C2I6 2025).

• Python
• PyTorch
• YOLOv8n
• LSTM
• CUDA (training)