AI AGENTS · LLM APPLICATIONS · MEASURED SYSTEMS

I build the whole path.
Then show where it breaks.

我把整条链路做通，也把它会在哪里失效讲清楚。

99.05%task successtask success
$35.68measured spendmeasured spend
0upstream 5xx @ 3×upstream 5xx @ 3×
10failure classes drilled已演练故障类别

Open to: AI agent & LLM application engineering · backend & distributed systems · data engineering & analytics校招方向：AI Agent 与大模型应用工程 / 后端与分布式系统 / 数据工程与分析

GitHubLinkedInEmail邮箱Phone电话

QWEN3.5-4B · RTX 4090 · $35.68 MEASURED · SHA-256 GATED

I don't just cite the number.
I open the command that made it.

Scaling free rule labels from 1,450 to 20,000 lifted a 4B model from 66.35% to 99.05% task success on complaint triage (n=2000, paired 95% CI), served with vLLM behind a C++20 token-aware gateway, scaled 1→3→1 on k3s. Total measured spend: $35.68. Under 3× overload the gateway sheds load with HTTP 429 and zero upstream 5xx; bare vLLM crashed at 5×. Distillation lost 14.2 pp to free rule labels and GRPO's CI includes zero — both runs are kept on the page.

NEGATIVE RESULT

Distilled SFT −14.2 pp vs rule SFT — complete-negative
OPEN THE RELEASE CONSOLE →
99.05% task success99.05%
n · CI
2000 · [98.60%, 99.45%]
Command
make reproduce-headline
SHA-256
sha256:9e547b3418ce1f633914e8dfe7b818fa6f7e064a891cee8e4c73837e6ce45f4d
0 upstream 5xx @ 3×0
n · CI
492
Command
make reproduce-headline
SHA-256
sha256:9e547b3418ce1f633914e8dfe7b818fa6f7e064a891cee8e4c73837e6ce45f4d
distilled SFT −14.2 pp vs rule SFT−14.2 pp
Command
make reproduce-headline
SHA-256
sha256:9e547b3418ce1f633914e8dfe7b818fa6f7e064a891cee8e4c73837e6ce45f4d

13-NODE GRAPH · TOOL GUARDS · ONNX INT8 · ON-DEVICE OCR

Agents that act.
Systems built to be second-guessed.

  1. Release Guardian132 live runs · 8/8 gates · 30/44 strict · citation fidelity 100%
  2. Triage Router+0.037 macro-F1 · −$120.58 / 1k calls · drift 2015–2026
  3. Privacy Preflight96 worker tests · OCR 19/19 hits · 2 FP · local-only

C++20 · GO · KAFKA · FLINK · K3S · POSTGRES

The agents sit on a stack.
The stack has its own receipts.

I break pipelines on purpose, then check recovery against source, table, and event state.

Gateway

Frontier Forge

0 upstream 5xx @ 3× · n=492

Serving

Frontier Forge

GPTQ-int4 p95 0.963 s @ 4 QPS

Stream

Exactly-Once Drills

10 failure classes drilled

Storage

Exactly-Once Drills · Crossover Study

0 snapshot diffs across 5 recovery drills

Orchestration

Release Guardian

132 funded live graph runs

5 RUNS THAT DID NOT WORK

Five runs.
All of them stayed on the page.

ConclusionnDispositionReceipt
Distilled SFT −14.2 pp vs rule SFTcomplete-negativeRECEIPT
GRPO +0.25 pp in both completed seeds; both CIs contain zero2 × 2000; aborted seed n=0The unchanged ten-step reward-variance guard stopped R4 v2; no three-seed aggregate or missing paired delta is fabricated.RECEIPT
30 of 44 strict all-trials residuals44Aggregate gates pass; 30/44 scenarios fail the strict all-trials outcome view. This is not a claim that all scenarios pass.RECEIPT
Amazon arm: null228153The Amazon null is scoped to this catalog, temporal split, five-core population, and observed history depths.RECEIPT
One-pass structured output: 0% task success48all original requests; uncovered two-pass rows count as failure; two-pass task success 100%RECEIPT

3 SECONDARY · 2 ARCHIVED

Not every project needs the front room.

Crossover Study43.9M reviews ingested · 15.5M five-core interactions · crossover: not foundRAG Quality Lab4/12 questions regressed · 11,309 docs · 130 enterprise questionsAsk PortfolioGroupConv Atlas80 SHAPES · 10 × 30 SAMPLES · GRAPH / EAGERMargin Control Tower15,809 Olist aggregate rows · 99,441 source orders · 10 fail-closed contract checksCredit Policy Desk120,000 scored loans · 24,000 later backtest rows · capacity-gated policy audit

HOW THIS SITE IS BUILT AND CHECKED

Every claim on this site
opens the same way this one does.

What was verified
The site's headline claims trace to preserved project artifacts and their file identities.
Evidence class
Recorded experiments, deterministic replays and fictional fixtures, identified separately by project.
Boundary
A recorded check describes its captured checkpoint. It does not establish a fresh run or production performance.
Files, hashes and methods
release.jsonLucisZhang/frontier-forge · 06de6e5c1d02
sha256:9e547b3418ce1f633914e8dfe7b818fa6f7e064a891cee8e4c73837e6ce45f4d
EOD manifestLucisZhang/portfolio-site · acf05ae78859
sha256:7867fb1ec1f91787ada2bb9141426b29273324f8d84b3eac6968572085120d32
Privacy manifestLucisZhang/portfolio-site · acf05ae78859
sha256:cfa56cd3ba453227863bb997c937bfcb41cd6e01e339cf806ac00689eac3aebe
Content reviewed
2026-09-06
Gate status
R2 source verification passed: 38 verified, 16 skipped, 54 total.
GitHubEmail

Content reviewed 2026-09-06