← Back to Home
Industry Solution — Industrial Manufacturing

Smart Manufacturing: Self-Evolving Optimization + Vision LLM

Two engines, one factory. Famou — Baidu's commercially available self-evolving super agent — writes and continuously improves the algorithms that decide schedules, cutting layouts, forecasts and simulations. Yijian — the vision LLM platform — turns every camera into an inspector that reads gauges, checks compliance and never files paper. Field-proven in production: 96.4% material utilization, +6.31% scheduling throughput, +20pp forecast accuracy, 2,700+ stores under vision AI.

Solution Panorama 01 SEE → 06 SCALE
96.4%
steel nesting utilization, 51/51 plates placed
+6.31%
injection-molding throughput, same 82 machines
+20pp
forecast accuracy vs human judgment (M+1–M+6)
95%+
average accuracy of Yijian chain models in production
2,700+
stores under Yijian vision AI (1,800+ tea, 900+ coffee)
3×SOTA
Famou tops ALE / MLE / Kernel Bench leaderboards
FAMOU AGENT YIJIAN VISION DUMU EDGE CLOUD-EDGE SYNC hours to minutes CFD · 1-week vision cold start · 3-6 months to 1 week
The Story — One Factory, Full Chain

Follow Hanwell Precision from paper and spreadsheets to a self-evolving factory

Instead of a feature list, here is one continuous story: a precision-parts manufacturer in Dongguan. Every step below is a real capability in this solution — with the production metric it moved. The interactive demos let you play each step yourself.

1
CHALLENGE — WHERE THEY STARTED
A manufacturer's day is four quiet battles
06:30 — a quality inspector photographs 200 paper records before shift start; the data reaches a spreadsheet by Friday. 09:00 — a planner fights 404 orders across 82 machines in a spreadsheet, schedules take days and still idle capacity. 14:00 — one aerodynamic design iteration means hours of CFD compute. 18:00 — next quarter's demand is a judgment call entered by hand. The losses are already on the schedule, long before anyone says "AI transformation".
4battles fought daily, all manual
2
STEP 01 — SEE
Every camera becomes an inspector; every gauge reads itself
Existing CCTV stays; a Dumu AI box taps the NVR streams at the edge, and the Yijian vision LLM platform runs the skills in the cloud. One prompt defines a skill — "detect the LCD gauge and read the value" — and the multimodal model cold-starts it: prompt → v1 skill → auto data collection → finetuned v2 at ~90% accuracy, in about a week instead of the 3–6 months small-model projects used to take. PPE violations, zone intrusion, foreign objects and spill events stream into one dashboard with evidence frames.
Yijian PlatformMultimodal L0Dumu AI Box
1 weekvision skill cold start (was 3-6 months)
90%v2 accuracy after auto data loop
3
STEP 02 — CUT
Silicon-steel nesting at 96.4% utilization
Transformer-core production needs 51 sub-plates cut from a catalog of 50 mother-coil widths. Famou evolves its own cutting algorithm — ALNS with knapsack repair plus simulated annealing — without enumerating every cutting pattern. All 51 plates are laid out at 96.4% material utilization, with fewer blade changes and fewer mother coils. Every point of utilization is raw material bought back.
Famou EvolveALNS + SA
96.4%utilization, 51/51 placed
4
STEP 03 — SCHEDULE
Same 82 machines, 21 more orders completed every month
404 tasks, 82 injection machines, hard machine-mold combination rules, backward scheduling from delivery deadlines. Manual experience planning completes 333 tasks. Famou's evolved coarse-to-fine scheduler — scored combos → ALNS + simulated annealing → fine scheduling → strict/relaxed execution — completes 354 tasks on identical capacity: +6.31% throughput, roughly 21 extra finished orders a month.
Coarse-to-FineALNS + SA
333→354tasks per month, same machines
+6.31%throughput gain
5
STEP 04 — SIMULATE
Wind-drag CFD drops from hours to minutes
Famou evolves a surrogate model on the customer's own vehicle mesh — GraphUNet with SAGEConv graph convolutions, TopKPooling, and custom losses (node MAE + pressure-smoothing + gradient consistency). Prediction quality holds while inference falls from hours to minutes. Designers ship more iterations per week with the same team; the aero roadmap stops waiting on compute.
GraphUNetSAGEConvTopKPooling
h→minsolve time, no accuracy loss
6
STEP 05 — FORECAST
Forecasts 20 points more accurate than human judgment
An automotive-parts supplier forecasts by customer × product, M+1 to M+6. Salespeople used to blend customer numbers with personal judgment — average deviation 45–65%. Famou mines 24 feature families (market, sales, opportunity, operations) and cuts deviation to 30–49%: +20pp accuracy versus human judgment, +8pp versus hand-built algorithm models. Operations finally plans against numbers it can trust.
24 Feature FamiliesM+1–M+6
+20ppaccuracy vs human judgment
+8ppvs hand-built models
7
STEP 06 — SCALE
The loop keeps evolving — and it compounds
Digitized cameras feed cleaner data; cleaner data makes Famou's models better; better plans make the vision alerts meaningful. The same evolution engine tops ALE Bench, MLE Bench and Kernel Bench — AI SOTA on public leaderboards — and ships as a commercial platform with pre-integrated DeepSeek-R1, ERNIE-5.0, Qwen3 and Kimi-K2. The story ends where it began: one factory, two engines, every battle won — and the system keeps improving while production runs." data-zh2="">
3× AI SOTACommercial PlatformContinuous Evolution
3×AI SOTA on public benches
24/7evolution while production runs

Engine 1 — Famou: The Self-Evolving Optimization Agent

Humans define the What; the LLM evolves the How — and keeps evolving while production runs

01How Famou Works — The Evolve Loop▶
Humans define the "What": evaluation criteria, initial solutions, optional background knowledge. The LLM evolves the "How": prompt sampling assembles tried code plus human ideas, LLM code generation writes candidate algorithms, an evaluator pool scores every candidate automatically, and the evolution code database keeps the elites — mutation and crossover breed the next generation. Simulated species evolution, applied to production algorithms.
23%faster Gemini-architecture kernel, evolved
~75%math problems at best-known solution
02Steel Cutting & Nesting Optimization▶
Transformer-core silicon-steel slitting: 51 target sub-plates cut from 50 mother-coil widths. Famou self-evolved an ALNS + knapsack-repair + simulated-annealing cutting algorithm that never enumerates all patterns — it keeps approaching high-utilization, low-coil-count executable solutions. Result: full layout for all 51 sub-plates at 96.4% utilization, minimizing scrap area, blade changes and mother-coil count. Targets any 2D nesting / bin-packing problem: sheet metal, glass, textile, leather.
96.4%material utilization
51/51plates placed, production-ready
03Production Scheduling Under Hard Constraints▶
Injection-molding scheduling: 404 monthly tasks, 82 machines, strong machine-mold combination rules (not every mold fits every machine; one mold per machine at a time), multi-mold parallel tasks, splittable tasks, backward scheduling from delivery deadlines. Famou's coarse-to-fine algorithm: build_scored_combos() initializes scored combos → ALNS + simulated annealing deep-optimize → fine scheduling layer → strict single/multi-machine execution with relaxed fallback. Completed tasks rose from 333 (manual) to 354 — +6.31% on identical capacity, ~21 extra orders a month.
333→354tasks on the same 82 machines
+6.31%throughput, zero new capacity
04Engineering Simulation Speed-Up▶
Automotive wind-drag CFD replaced by an evolved surrogate: GraphUNet backbone with SAGEConv graph convolutions on the customer's vehicle mesh, TopKPooling to compress the graph, custom losses (node MAE + adjacent-node pressure-difference smoothing + gradient consistency). Accuracy holds; solve time falls from hours to minutes — changing how often designers iterate. The same evolution engine optimizes kernels and operators: +23% on a critical Gemini-architecture matrix-multiplication kernel.
h→minwind-drag solve time
23%kernel speed-up, evolved
05Demand Forecasting & Factor Mining▶
Automotive-parts demand planning by customer × product, M+1 through M+6. Famou mines 24 feature families — market data (auto sales, Tier-1 purchases, panel supply and share), sales data (customer forecasts, judgment, actuals, budgets, MTP), opportunity data (design-win and on-going pipeline, win rate, lost-deal lifecycle), operations data (panel/module capacity history, M+3+9 and Y+5 plans). Average deviation: human judgment 45-65%, hand-built models 40-54%, Famou 30-49% — +20pp vs humans, +8pp vs manual models.
+20ppaccuracy vs human judgment
24feature families mined
06Third-Party Proof: 3× AI SOTA▶
Same engine, different scoreboard. Famou tops ALE Bench (combinatorial optimization, benchmark published by Sakana AI from AtCoder problems), MLE Bench (ML engineering, published by OpenAI from Kaggle competitions) and Kernel Bench (performance optimization, published by Stanford) — AI state of the art on all three. Shipped as a commercial platform: management console, evolution engine, knowledge services (expert experience library, domain paper base, industry operator library) and pre-integrated DeepSeek-R1, ERNIE-5.0, Qwen3, Kimi-K2 plus API access to frontier international models.
ALEAI SOTA — combinatorial
MLEAI SOTA — ML engineering
KernelAI SOTA — performance
LIVE DEMO — Nesting Utilization Explorer
Move the sliders: demand mix and trim tightness — the nesting strip and utilization gap recompute instantly
SIMULATED NESTING MODEL
The strip below renders sub-plate layouts (cyan) vs scrap (red) — manual planner on top, Famou-evolved on the bottom.
MANUAL
89.2%
FAMOU
96.4%
LIVE DEMO — Scheduling Throughput Simulator
Set machines, order pool and constraint tightness — manual vs Famou-evolved completion recomputes live
SIMULATED SCHEDULER
Manual planning
333
Famou-evolved
354

Engine 2 — Yijian: The Vision LLM Platform

One prompt defines a vision skill; the multimodal LLM cold-starts it in a week — cameras, gauges, compliance, all in one platform

01Platform Architecture — Cloud × Edge × Models▶
Yijian public-cloud platform on top of a vision-LLM base: CV large models + multimodal L0 model, full video-network data access, scenario model factory with a complete toolchain — data backflow, smart filtering, assisted labeling, quality diagnosis, evaluation analytics, skill orchestration, CV-model re-checking and multimodal finetuning. Below it, Dumu edge hardware (AI boxes and AI cameras) runs skills in-store or in-plant; above it, industry model libraries (F&B, retail, chain stores) and native apps: inspection tours, duty audits, material counting, shelf compliance, operations analytics.
Cloudskill mgmt + device mgmt + events
EdgeDumu boxes & cameras, second-level alerts
02Multimodal L0 — Vision Tasks From Natural Language▶
The multimodal L0 model takes images plus natural language, aligns visual and text features, and autoregressively outputs answers — VQA, grounding, captioning, detection, segmentation from one model. It reads an LCD scale in plain language ("a scale with a paper cup on it, reading 325.0 g"), handles complex semantics like REC, and finetunes fast: 3,900 samples finetuned in 30 minutes. In the small-model era every task needed a custom pipeline (LCD detector → OCR → regex), 3-6 months to first delivery and 2-week iteration cycles; with the L0, one prompt tunes a skill in 1-2 days — One-For-All tasks, better generalization on long-tail scenarios.
30 minfinetune on 3,900 samples
1-2 daysper prompt tuning cycle (was 2 weeks)
03Cold Start in 10 Steps — New Skill in About a Week▶
For long-tail, fragmented scenarios with few or zero samples: 1 analyze the requirement → 2 define the scenario → 3 decompose into an explicit prompt ("detect the LCD screen, read the value") → 4 connect devices at scale → 5 orchestrate and launch v1 on the platform → 6-7 collect and auto-label live inference data → 8 finetune the large model → 9 ship v2 at ~90% accuracy → 10 keep iterating. The whole loop runs in about a week, then the AIGC-assisted data engine keeps compounding accuracy.
~1 weekfrom prompt to 90% v2 skill
Autodata collection + labeling loop
04Dumu Edge Hardware — Boxes & Algorithm-Defined Cameras▶
Two deployment routes. Retrofit: keep existing cameras, add a Dumu AI box or frame-grab box next to the NVR — CV186 handles 4-8 channels, BM1688 8-10, BM1684x up to 32 concurrent streams, 1-3 AI skills per channel, average 95%+ accuracy across chain-store model projects, edge inference for low latency and local data processing. New-build: Dumu AI cameras are algorithm-defined — VOF-S / DC-PL2 / DC-PD/PM families, 2 to 4.8 TOPS, dark-light full-color, fixed or zoom, face-capture or long-focus, private or subscription licensing. Both routes connect to the Yijian platform for skill push-down, event reporting and data backflow for model iteration.
4-32channels per box
95%+avg accuracy in production projects
05Proven at Scale — 2,700+ Stores, 12 Production Skills▶
Chain tea franchise, 1,800+ stores, cloud-side analysis: trash-bin compliance (open / overflowing / items on lid), fridge door left open while unattended, table foreign objects (cups, straws, napkins after guests leave), items placed on ice machines (heat risk), staff uniform and hairnet compliance — each event carries store ID, device ID, timestamps and evidence frames into a food-safety risk score per store. Chain coffee brand, 900+ stores, edge-side analysis adds: unauthorized entry during closed hours (22:00-07:00), staff absence from the prep zone >5 min, video-quality faults, material-consumption counting (cup / lid tearing). Same skill engine transfers directly to plants: PPE, zone intrusion, machine-state and gauge reading, spill and foreign-object detection.
1,800+tea stores, cloud-side
900+coffee stores, edge-side
12production skills live
06Industrial Scenario Pack▶
The skills already proven in chains map one-to-one onto plant floors: uniform/hairnet compliance → PPE detection (helmet, vest, goggles); closed-hours intrusion → hazardous-zone access control with person-type classification (operator / logistics / outsider); table foreign objects → foreign-object and spill detection on lines and walkways; fridge-door state → cabinet, valve and gate states; LCD scale reading → gauge, dial and HMI meter reading via multimodal VQA; staff absence → unmanned-station alerts; material counting → consumable and WIP counting. Skills combine with Famou: vision events become structured inputs, Famou's planners and forecasters consume them, and the factory closes the loop from seeing to deciding.
PPEhelmet, vest, goggles
GaugeLCD / dial / HMI reading
Loopvision feeds Famou planners
LIVE DEMO — Vision Skill Runner
Pick a skill prompt: the simulated camera frame scans, detects, reads — events stream to the log
SIMULATED MULTIMODAL INFERENCE
CAM-07 · LINE-2 · 1080P · LIVE
Select a skill prompt above — inference runs on this simulated frame.
LIVE DEMO — Edge Deployment Planner
Enter sites and cameras — the planner picks retrofit vs new-build and sizes the hardware
DEPLOYMENT CALCULATOR

Verify It Yourself — Test Cases

TRY = run it live on this page right now. PASS = verified production result.

TRY ITTC-01Famou's nesting holds a utilization gap as orders diversify▶
GivenThe Nesting Explorer demo above, defaults: 5 order types, medium trim, 30 coil widths
WhenYou drag order mix diversity to 12 and trim tolerance to loose
ThenBoth strips re-render; Famou stays in the 94–97% band while manual drifts down — the gap widens exactly where the real case found it
ProvesProduction behavior: 51/51 sub-plates at 96.4% on real silicon-steel orders
TRY ITTC-02The scheduler finds +6% capacity at reference settings▶
GivenThe Scheduling Simulator at the real case's settings: 82 machines, 404 tasks, tightness 0.6
WhenYou ease constraint tightness to 0.2 — or raise the pool to 600
ThenBoth bars grow, but the Famou bar keeps a 6%+ lead; the verdict converts the gap into extra completed orders per month
ProvesProduction result: 333 → 354 tasks (+6.31%) on identical capacity, ~21 extra orders/month
TRY ITTC-03One prompt cold-starts a vision skill and reads a gauge▶
GivenThe Vision Skill Runner above, simulated camera CAM-07 on LINE-2
WhenYou click "Detect the LCD gauge and read the value", then try the other three prompts
ThenThe scanline sweeps, bounding boxes pop with labels and confidence, the VQA answer streams to the event log with timestamps — same event schema the production platform emits
ProvesPlatform behavior: prompt → v1 skill → data loop → ~90% v2 in about a week (vs 3-6 months small-model delivery)
TRY ITTC-04The edge planner sizes hardware for 500 sites without over-provisioning▶
GivenThe Edge Deployment Planner, retrofit mode, 120 sites × 4 cameras
WhenYou push sites to 500 and cameras to 16, then switch to new-build
ThenThe planner re-selects the box model for the channel count (CV186 / BM1688 / BM1684x, 4–32 channels) or switches to AI cameras, and shows the cloud-sync architecture
ProvesHardware line card: 4-32 concurrent streams per box, 1-3 skills per channel, 95%+ average project accuracy
PASSTC-05Silicon-steel nesting: 96.4% utilization, 51/51 plates placed▶
SetupTransformer-core production; 51 sub-plates from 50 mother-coil widths; evolved ALNS + knapsack repair + simulated annealing
ResultFull executable layout for all 51 sub-plates at 96.4% utilization, minimizing scrap area, blade changes and mother coils
PASSTC-06Injection scheduling: 333 → 354 tasks, +6.31% on same capacity▶
SetupOne month of real orders: 404 tasks, 82 machines, machine-mold combination constraints, deadline backward scheduling
ResultCoarse-to-fine evolved scheduler completes 354 vs 333 manual — +6.31% throughput, ~21 extra orders/month, schedule exported to the line
PASSTC-07Demand forecasting: +20pp accuracy vs human judgment across M+1–M+6▶
SetupAutomotive-parts supplier; 10 customer-product combinations; 24 feature families across market / sales / opportunity / operations data
ResultAverage deviation: human judgment 45-65%, algorithm models 40-54%, Famou 30-49% — +20pp vs humans, +8pp vs manual models
PASSTC-08Yijian vision platform runs 2,700+ stores with 12 production skills▶
SetupChain tea franchise 1,800+ stores (cloud-side: trash bin, fridge door, table objects, ice machine, uniform) + chain coffee 900+ stores (edge-side: + intrusion, absence, video quality, material counting)
Result95%+ average accuracy across model projects; per-store food-safety risk scores with evidence frames; 4-32 channels per edge box
PASSTC-09Wind-drag surrogate: hours to minutes with no accuracy trade-off▶
SetupCustomer vehicle mesh data; evolved GraphUNet + SAGEConv + TopKPooling with node-MAE / pressure-smoothing / gradient-consistency losses
ResultPrediction quality holds while inference drops from hours to minutes per design iteration

Start with a 3-Move POC

Move 1 — Discovery workshop (2 weeks): map 2-3 highest-value problems, check data, agree baselines. Move 2 — Proof of concept (4-6 weeks): Famou evolves against your historical orders while Yijian cold-starts vision skills on your cameras. Move 3 — Scale & evolve: MES/ERP integration, ops handover, continuous evolution cycles. Bring one problem owner, sample data, and a baseline you trust.

Contact Us Partner Program