Career
R&D Engineer — InsightOS, Shanghai R&D Engineer — InsightOS
Robot media I/O, sensor synchronization, and cabinet-operation tooling.
01 / 09 — Profile
Yixin Zhang
R&D engineer building dependable sensing and control infrastructure.
My goal
Be challenged in robot learning—and shorten the loop from failure to a better policy. Robot learning: turn deployment failures into better policies.
Industrial robot programs can contribute fast iteration, compute, operational data, and real failure cases; school research can contribute depth, open evaluation, and room to test new ideas. I want the two to change each other, not an application-only track centered on delivery.
Dedicated to extreme ownership and proactive problem-solving—taking one extra step when it improves the result.
Career
Industrial Digitalization Engineer — Beijing Zhongke Shengtai Industrial Digitalization — Zhongke Shengtai
Industrial-wastewater vision PoCs and a floc-analysis handoff.
Career
Scientist Intern — SIMTech, A*STAR Scientist Intern — SIMTech
ASEAN energy-system studies converted into model-ready data.
Volunteering
SG Beach Warriors
Coastal cleanup across Singapore’s north and east shores.
Education
NUS-ISS — Graduate Diploma in Systems Analysis NUS-ISS — Systems Analysis — overlaps HIT
GPA 4.61/5.0 — online Jun 2023–Jun 2024; in person Jun 2024–Mar 2025.
HIT project
DQN path-planning capstone
Delayed observations, reward shaping, and dynamic-obstacle simulation.
HIT project
SparkDesk KBQA & security evaluation
Agricultural RAG, 100+ questions, and prompt-injection tests.
HIT-period project
LLM-Cookbook contributor
Terminology, localized prompts, reproducibility checks, and demos.
Social practice
Exploring the Pillars of a Great Power
HIT field-research team lead — second prize and key incubation team.
HIT project
Video-text restoration & recognition
First prize; selected for national-level continuation.
Education
HIT — B.Eng., Intelligent Test & Control Engineering HIT — B.Eng., Intelligent Test & Control
Electrical Engineering — 86.95/100 — People’s Scholarship ×5.
02 / 09 — Current work
Closing the cabinet-operation loop
A reusable robot media runtime underneath perception, pose estimation, and contact.
R&D Engineer — Sep 2025–present
上海具识智能有限公司 — InsightOS
Problem
Keep sensing stable while the hardware changes
Cameras and depth sensors move across devices; open-loop motion becomes brittle when target and robot frames drift.
Built
- Media I/O daemon, memfd IPC, ROS2 bridge, C++ SDK, and
insightos://routing. - Device discovery, topology-drift recovery, and cross-view time synchronization.
- YOLO11 segmentation, RGB-D pose estimation, ROS TF transforms, and visual/tactile correction.
C++ — Linux — ROS2 — DDS — RGB-D — shared memory — REST/CLI
Needs review: the video shows physical door and breaker contact, not electrical state change, the internal pipeline, success rate, or production reliability. The résumé’s “>95% within 15 minutes” metric still needs logs and a denominator; confirm employer clearance for internal-cabinet frames.
Needs review: the frames show physical contact—not electrical state, autonomy, or reliability. Internal-cabinet publication clearance is still required.
click to load — hosted video failed; use the direct link.
03 / 09 — Industrial computer vision
From floc video to a binary operator cue
Implemented evidence, reported evaluation, and the next temporal study are kept distinct.
Industrial Digitalization Engineer — 2025
Beijing Zhongke Shengtai
Problem / decision
Make an operator’s visual judgement inspectable
Underwater frames are noisy and flocs overlap. The desktop PoC exposes the source frame, binary mask, thresholds, features, and a binary cue: coagulation sufficient or add more coagulant.
2,270 labeled images
258,887 extracted floc records
Implemented method
archived code
grayscale→median blur→adaptive threshold→open / close / erode→components→XGBoost
- Count
- population of valid connected components
- Major-axis diameter
- elongated floc scale
- Equivalent diameter
- area expressed as a circle
- Circularity
- shape compactness, 4πA/P²
- Area fraction
- scene coverage by retained flocs
The features were chosen to cover population, scale, shape, and coverage with inexpensive geometry. The archived classifier is binary, not a five-step dosing controller.
Experiment / reported result
PoC evidence
88.5%
XGBoost test accuracy
87.9%
XGBoost weighted F1
67.4%
fixed-threshold accuracy
| Classifier | Accuracy | Recall | F1 |
|---|---|---|---|
| XGBoost | 88.5% | 87.3% | 87.9% |
| Threshold | 67.4% | 90.2% | 77.2% |
Needs review: image labels were replicated to individual flocs, then randomly split. Video/day/plant independence and a held-out future period were not tested; 89% is the rounded résumé value.
Research extension
reference results, not my experimentPearson screening is useful for the next study only when redundancy and dose relevance are both visible.
Floc morphology ↔ morphology
area / circumference0.72
area / circularity0.70
fractal dimension / circularity0.71
High pairwise values flag redundant descriptors and unstable attribution.
Floc morphology ↔ coagulant dose
area0.61
circumference0.58
fractal dimension0.55
circularity0.60
Association supports screening, not causality or a dosing optimum.
Causal window protocol
frames→1-min robust summaries→x[t−W:t]→plant lag τ→y[t+τ:t+τ+H]
Estimate τ from HRT and training-only cross-correlation. Compare W = 2.5 h, 4 h, and 24 h as reference candidates; split by time/video/day and purge at least W+τ+H between folds.
A second supplied study is directionally similar, not numerically transferable
count 0.883
area ratio 0.745
equivalent diameter 0.921
fractal dimension 0.864
PAC-dose correlations from the blue conference slide. Different variables and study conditions explain why these coefficients should not be merged with the matrix above.
Reference only: I used these findings but did not conduct the reported experiments. Correlation screening must be recomputed inside each training fold. TCN/Transformer and SHAP do not appear in the archived PoC.
04 / 09 — Climate analytics
Turning case studies into model-ready data
Data contracts connect published assumptions to optimization and country-level mitigation choices.
Scientist Intern — Sep 2024–Jan 2025
SIMTech, A*STAR — Singapore
Platform / problem
public contextGlobal Mitigation Potential Atlas
The Atlas is a decision-support platform for exploring country mitigation potential at different costs and how collaboration can unlock additional strategies.
Its pilot methodology focuses on technical potential at cost levels, translating national systems into comparable, optimization-ready cases.
My contribution
- Mapped ASEAN case-study prose, equations, assets, constraints, trade links, units, and CO₂ budgets into consistent JSON.
- Built a schema-aware, LLM-assisted conversion workflow and inspected optimization output against prior OSeMOSYS/TIMES cases.
- Contributed to early MOYA Cascade concepts; the launched product and platform modeling were broader team efforts.
Needs review: “Pass@1 >85%” lacks its dataset, rubric, sample count, and failures. Publication clearance is still required for project-only scenario crops.
Method / data contract
internship artifacts
Experiment / scenario comparison
modeled outputWhat changes when Singapore and Indonesia are modeled as isolated systems versus an idealized collaboration?
These are model outputs used to inspect scenario behavior, not observed forecasts. My contribution was data translation and workflow prototyping, not ownership of the Atlas or its full quantitative model.
05 / 09 — HIT capstone
Path planning with delayed observations
Discrete DQN actions, delayed state estimates, and repeated simulation comparisons.
B.Eng. capstone — 2023–24
HIT — Electrical Engineering
Problem / definition
Plan while the newest observation is already stale
A dynamic simulator injects observation delay while the agent must still reach the goal and avoid moving obstacles.
Method
delayed state s
- Finite action candidates are confirmed by the preserved argmax decision rule.
- Replay memory, main/target networks, and randomized observation delay.
- Reward terms cover progress, collision avoidance, and convergence.
Needs review: the exact action catalogue, raw logs, seeds, confidence intervals, simulator configuration, and real-robot evaluation are unavailable.
Simulation
complete figure
Learning curves
supplied artifact
Experiment / reported mean
15 repetitionsProposed
98.1%
Without reward logic
97.2%
Traditional DQN
92.6%
Traditional Q-learning
93.1%
| Method | Success | Mean path length |
|---|---|---|
| Proposed | 98.1% | 29.113 |
| Without reward logic | 97.2% | 29.474 |
| Traditional DQN | 92.6% | 30.852 |
| Traditional Q-learning | 93.1% | 32.224 |
98.1% versus 92.6% is +5.5 percentage points (+5.9% relative). Values are transcribed from the capstone slide; raw logs are unavailable.
+5.5 pp success versus traditional DQN
−1.739 mean path length versus traditional DQN
−0.9 pp after ablating the proposed reward logic
06 / 09 — HIT innovation project
Recognizing text across blurred frames
A quality-gated multi-frame path retains usable history, fuses evidence, and recognizes a sequence.
Team member — 2021–22
Innovation & Entrepreneurship Programme
Problem / contribution
Recover strokes that a single blurred frame cannot preserve
The team system detected and matched text instances, retained the best adjacent observations, fused aligned features, and recognized the sequence.
Documented role: Transformer-inspired recognition and coupling recognition features to the upstream restoration network.
First prize at HIT; selected for national-level continuation.
Needs review: the report supports recognizer/feature coupling, not the broader résumé deblurring scope, Android deployment, or “+0.23% F1” wording.
Original quality-gated pipeline — redrawn
team algorithm
Incoming text croptracked instance at frame t
→
Quality scorecompare with matched history
→
In TOP5?quality gate
YES
Current crop = referenceupdate TOP5 quality memory
In TOP4?update four saved support crops
NO
Keep prior outputdo not rerun the recognizer on a worse crop
support 1support 2support 3support 4reference t
→Feature extractionper crop
→Align + fusemulti-frame evidence
→SVTR-inspired recognizermixing blocks + CTC
→Text sequencelatest accepted output
Faithful to the project flow: only a sufficiently good matched crop becomes the new reference; four saved support crops and the reference are fused before recognition.
Experiment definition
- Because no public blurred-video corpus was available, blur kernels were applied to ground-truth still images to form consecutive degraded frames.
- Chinese Benchmark and ICDAR 2019-LSVT were used for the reported recognition comparison.
- The final recognizer is SVTR-inspired with local/global mixing blocks and CTC prediction.
Exact sample counts, blur distribution, seeds, and raw predictions are not preserved.
Reported recognition accuracy
acceptance table
69.2%
Chinese Benchmark
68.5%
ICDAR 2019-LSVT
| Method | Chinese | LSVT |
|---|---|---|
| Team method | 69.2% | 68.5% |
| SVTR-S | 69.0% | 68.2% |
The table is preserved evidence; raw predictions and an independent rerun were not supplied.
Perceivable qualitative outputs
team slide
→
我是第一段文字哈哈哈哈
→
我是第三段文字呀呀呀
07 / 09 — Open work
Public code and model evaluation
Two merged vector-search changes, an agricultural KBQA assessment, and early prompt-engineering coursework.
Inspect the artifacts
Public links wherever preserved
camel-ai/camel — PR #3758
Feb 2026 — merged
Feb 2026 — merged
Qdrant query filters
Added distance metrics and filtering, with tests and examples around the new query behavior.
labring/FastGPT — PR #6348
Jan 2026 — merged
Jan 2026 — merged
OceanBase HNSW quantization
Added SQ/BQ vector-index quantization levels and corresponding distance constraints.
SparkDesk KBQA
Oct—Nov 2023
Oct—Nov 2023
Agricultural KBQA assessment
RAG knowledge base, closed/open questions, prompt injection, and jailbreak tests; résumé reports 100+ questions and 10+ attacks.
datawhalechina/llm-cookbook
2023
2023
Prompt-engineering coursework
Proofreading, terminology/API alignment, prompt reproduction, and chatbot experiments. Exact contribution commits still need linking.
Smaller exploration:
decentralized agent identity and E2EE messaging through the ANP handle zyx.awiki.ai.
08 / 09 — Credentials
ML, cloud, and data credentials
Seven professional certifications and two NVIDIA DLI course certificates visible in the supplied record.
Issued Dec 2024—Jan 2026
Expiry shown where supplied
NVIDIA
1 professional2 DLI courses
- Certified Professional: Gen AI LLMs
- Disaster Risk Monitoring Using Satellite Imagery
- Introduction to Transformer-Based NLP
Oracle
3 professional- OCI 2025 DevOps Professional
- OCI 2025 Data Science Professional
- OCI 2025 Generative AI Professional
Microsoft
2 associate- Azure Data Scientist Associate
- Power BI Data Analyst Associate
AWS
1 associate- Machine Learning Engineer — Associate
09 / 09 — Questions
What is worth testing next?
Systems engineering now; robot learning next. Keep the questions close to physical evidence.
Working principle
Make assumptions observable and falsifiable
Questions I’m chasing
Turn robot failures into better experiments
- How can a robot verify that a physical task is actually complete?
- How should perception, calibration, and control error be represented and recovered?
- Which moments—contact, failure, intervention—are worth retaining or sharing across robots?
- Which assumptions in LLM-driven robotics can be formalized, checked, and falsified?
What keeps me curious
Systems, ideas, and fieldwork
- Inference-time learning, world models, action-conditioned models, and long-horizon memory.
- Knowledge workflows that surface conflicts instead of silently flattening them.
- Science translation, sustained reading, and coastal volunteering outside engineering work.
Decentralized multi-agent infrastructure: a practical touchpoint is the ANP handle zyx.awiki.ai.
AI + formalized mathematics: still an early interest; current foundations are formal languages, context-free grammars, and propositional logic.
←→ navigate
Slide 1 of 9: Profile