Syder 9 Solutions · syder9solutions.xyz

Syed Ali Haider

Systems Analyst & Technical Solutions Specialist | Business Systems Analysis · API & Workflow Automation

Karachi, Pakistan. Available for remote work worldwide.

I build AI and backend systems that ship with test suites, calibrated evaluations, and numbers you can reproduce. Twelve of them are below. Every metric links back to a public repository with the code that produced it.

Proof, not adjectives

Twelve systems, four clusters

Every card shows the measured headline result from the project's own evaluation report. Filter by cluster or read them all.

RAG & Retrieval

chat-with-your-docs-rag

Production RAG engine with hybrid BM25Plus + dense retrieval, RRF fusion, cross-encoder reranking, clickable citations, and a strict refusal guardrail.

Measured: 100% refusal accuracy on the adversarial set, Recall@5 1.000, judge agreement kappa = 0.857, p95 8.6 ms local.

  • FastAPI
  • BM25Plus
  • RRF
  • Cross-Encoder
  • Citations
View repository: chat-with-your-docs-rag

semantic-search-engine

Retrieval-only search engine over BEIR trec-covid: dense FAISS + BM25 hybrid retrieval, RRF fusion, and cross-encoder reranking with per-stage IR evaluation.

Measured: hybrid + reranker nDCG@10 0.808 and precision@5 0.940 (dense baseline 0.785 / 0.908); dense MRR 0.970.

  • FAISS
  • BM25
  • RRF
  • Cross-Encoder
  • BEIR
View repository: semantic-search-engine

multimodal-doc-ai

Vision-LLM document intelligence over charts, tables, and invoices: visual RAG from page pixels with strict judge evals against a text-only baseline.

Measured: visual-only QA accuracy 0.556 vs 0.056 for the text baseline; overall 0.775 vs 0.700 across 120 paired rows.

  • Vision LLM
  • Visual RAG
  • FastAPI
  • McNemar
  • Bootstrap CI
View repository: multimodal-doc-ai

LLM Engineering Depth

llm-from-scratch

Decoder-only GPT implemented from first principles in PyTorch: byte-level BPE, multi-head attention, training loop, and sampling. No shortcuts.

Measured: loads OpenAI's published GPT-2 (124M) weights and matches HuggingFace logits to 4.6e-05 max absolute difference, 100% argmax agreement.

  • PyTorch
  • BPE
  • Attention
  • KV Cache
  • GPT-2 Parity
View repository: llm-from-scratch

Production Discipline

production-eval-suite

Evaluation harness that gates deployments in CI: error-analysis golden set, calibrated LLM judge, and statistical significance testing on every prompt change.

Measured: judge kappa = 0.909; gate blocked a real regression (delta -0.724, 95% CI [-0.869, -0.576], p = 0.0078) via Wilcoxon + bootstrap CI.

  • LLM Judge
  • Wilcoxon
  • Bootstrap CI
  • GitHub Actions
View repository: production-eval-suite

structured-extraction-pipeline

Schema-locked invoice/receipt extraction with a bounded repair loop, per-field confidence scoring, human-review routing, and hallucination detection.

Measured: field accuracy 1.00, schema valid after repair 1.00, hallucination rate 0.00, review precision 1.00 on the hand-labeled eval set.

  • Pydantic
  • Structured Outputs
  • Repair Loop
  • FastAPI
View repository: structured-extraction-pipeline

guardrails-moderation

Defense-in-depth guardrails for an LLM chat endpoint: PII detection and redaction, prompt-injection defense, document fencing, and human-gated tools.

Measured: attack success rate 0.738 to 0.000 across 42 red-team attacks; 0/20 benign inputs over-blocked; guard cost p95 1.62 ms.

  • Presidio
  • Injection Defense
  • Red-Team Evals
  • HITL
View repository: guardrails-moderation

llmops-cost-dashboard

LLMOps observability: end-to-end call tracing, cost per request/feature/user, latency percentiles, exact + semantic caching, and regression alerts.

Measured: semantic cache reached 45.0% hit rate and cut p95 latency 60.3% (14694 ms to 5838 ms) on a live workload; cost/1k fell 84.4% (mixed model tiers, see repo caveat).

  • FastAPI
  • Semantic Cache
  • SimHash
  • SQLite
  • Alerting
View repository: llmops-cost-dashboard

Agents & Voice

agentic-workflow

OpsRescue: autonomous SRE incident-remediation agent with poka-yoke tool schemas, typed error semantics, retry budgets, and an independent verifier gate.

Measured: pass^3 trajectory reliability 100% (standard and fault-injection), transient fault recovery 100%, p95 turnaround 250.8 ms.

  • ReAct
  • Poka-Yoke
  • Pydantic
  • Trajectory Evals
View repository: agentic-workflow

mcp-second-brain

Production Model Context Protocol server exposing a sandboxed markdown knowledge vault over stdio and HTTP/SSE, with version-pinned tools as a trust boundary.

Measured: 13/13 red-team path-traversal and exfiltration attempts blocked, schema accuracy 8/8, hybrid retrieval Recall@3 100%, 33 tests passed.

  • MCP
  • stdio
  • HTTP/SSE
  • Sandboxing
View repository: mcp-second-brain

realtime-voice-agent

Streaming voice agent (STT to LLM with tools to TTS): sentence-level streaming, barge-in interruption, and per-stage latency profiling over WebSocket.

Measured: time-to-first-audio 550 ms p50 live against a real server; offline eval 8/8 pass; live end-to-end run verified tool calls and barge-in.

  • WebSocket
  • STT
  • Streaming TTS
  • VAD
  • Tool Calls
View repository: realtime-voice-agent

Experience & education

  1. Current

    Solutions & Automation Specialist

    Syder 9 Solutions · Self-employed · Remote

    Designing and shipping AI and backend systems end to end: specification, implementation, independent review, calibrated evaluation, and release. The twelve projects above are this work.

  2. Nov 2021 - Jun 2025

    Systems Analysis, Defect Triage, SOPs

    Sydney

    Business systems analysis, defect triage, and standard operating procedures: turning ambiguous requirements into testable specifications and repeatable processes.

  3. Aug 2020 - Sep 2021

    IT Operations

    Sydney

    Network and infrastructure operations: routers, switches, TCP/IP, DNS/DHCP. Keeping production systems reachable and diagnosable.

  4. Education

    Master of Business Information Systems

    Torrens University · Australia

How I work

Systems analysis first

My background is business systems analysis: requirements elicitation, defect triage, and SOPs. That discipline carries into engineering. Before code, I write down what the system must do, what failure looks like, and how success will be measured.

Then I build the smallest system that satisfies the spec, have it independently reviewed, and gate it on evaluations that were designed before the implementation existed.

Evidence-first engineering

  1. Spec. Write the contract: inputs, outputs, failure modes, acceptance metrics.
  2. Implement. Build against the spec with tests from day one.
  3. Independent review. A separate reviewer attacks the work: adversarial probes, red-team sets, falsifiability checks.
  4. Evals. Calibrated judges, statistical significance, honest baselines. Numbers that move when the system breaks.
  5. Commit. Ship only when the gates pass. Every published figure traces to a reproducible artifact.

If a metric cannot move when the system is broken, it certifies nothing. Every project above includes the harness that proves its numbers are real.

Contact

Open to remote roles and engagements worldwide. The fastest way to reach me is email.