谭新 Tec_anX

谭新 · Tec_anX

Pursue exceptional efficiency.

Explore deeply. Build systems ready for production.

I develop software, integrate algorithms and connect devices. In AISOP, I build the frontend, backend and SOP state machine. At Schaeffler, I lead interface coordination and field integration.

Independent judgement · Cross-domain delivery · Continuous growth

How I grow

Across disciplines. Into practice.

Own judgement

Software & architecture

I develop the AISOP frontend, backend and SOP state machine, connecting visual inference to production workflows and device interfaces.

01 / Projects

Systems I build. Problems I solve.

01Software depth

AISOP · Process decision platform

Connect visual recognition to process rules for step transitions, round completion and exceptions.

The problem

Recognition alone cannot establish that a step is complete. Decisions also depend on step order, current state and exception rules.

My responsibility

Develop the frontend, backend, configuration and data management; integrate inference and the SOP state machine; adapt scanners, status lights and MES interfaces; deploy and run regression checks.

Evidence and boundaries

Platform features are complete and in production. Specific field outcomes remain subject to runtime evidence and acceptance.

Inference & process decisionsStatic relationship diagram

Select a node to inspect; select it again to release.

Illustrative system relationships, not field data or acceptance results.
Explore decisions and architectureClose decisions and architecture

Implementation

Build result visualization, configuration and data management, then connect annotation and training. Define step decisions, transitions, round completion and exceptions before integrating devices and MES.

Trade-offs

Define states and transitions before device integration. Move inference from PyTorch to ONNX to reduce the release footprint; compare changes with the regression baseline.

Architecture

Capture → ONNX inference → SOP state machine → Records → Devices / MES

Review

Inference, UI and devices must follow the same process state; working components alone do not establish a working system.

Ask about the work ↓

02Integration breadth

Schaeffler · Device and protocol integration

Integrate MQTT, RS-485 lamps and multi-channel audio; troubleshoot motion inputs across devices and applications.

The problem

Multiple teams supply algorithms, devices and workstation apps. Input formats, interfaces and verification conditions must align.

My responsibility

Project lead for technical coordination; hands-on sample capture, interface adaptation and device troubleshooting; coordinate motion data, SOP and dashboard inputs.

Evidence and boundaries

Ongoing: tools and partial integration records exist. Synchronization, calibration, real-time accuracy and full-chain stability need independent checks and field acceptance. Core21 does not establish finger tracking. Voice control and moving projection remain pending or separate changes.

Communication & device responseStatic relationship diagram

Select a node to inspect; select it again to release.

Illustrative system relationships, not field data or acceptance results.
Explore decisions and architectureClose decisions and architecture

Team scope: colleagues develop calibration, motion reconstruction, SOP models and dashboards. Partners provide digital twin presentation and site installation.

Implementation

  • Organize and participate in calibration/action capture and handover; report re-calibration issues.
  • Adapt cross-network interfaces; report dropped frames and invalid motion inputs.
  • Diagnose RS-485 wiring, lamp mapping and switching timing; test MQTT, multiple speakers and microphones.
  • Collect workstation samples and align SOP inputs, dashboard data and deployment dependencies.

Trade-offs

Test inputs, communication and device feedback separately before combining them. Verify scope changes separately from the original plan.

Architecture

Detections → MQTT relay → RS-485 lamps → Warning feedback

Motion / Core21 → Cross-network interface → Digital twin (partners); audio linkage pending.

Review

Record interfaces, versions and test results after each site change. Keep runnable, connected and accepted states distinct.

Ask about the work ↓

Further practice

Web development, proposal automation and an unsuccessful Agent trial.

03

Company website · Content and development

Independently research product information and build multilingual pages and interactions.

Result & boundaryThe multilingual site and design documentation are complete. Indexing, generative citations and conversion still need data.

Explore decisions and architectureClose decisions and architecture

The problem

Visitors need a clear reading order to understand AI-SOP, related products and the platform.

My responsibility

Independently design and develop the site, confirming direction with the owner; organize product content and implement multilingual pages, SEO/GEO metadata and interactions.

Implementation

Verify product definitions and visitor questions, establish the reading order, then implement pages with narrow screens, language lengths and keyboard access in mind.

Trade-offs

Set content and product relationships before visuals and motion, so presentation supports a clear explanation.

Architecture

Source material → Product definitions and hierarchy → Multilingual pages and metadata → Responsive interactions

Evidence and boundaries

The multilingual site and design documentation are complete. Indexing, generative citations and conversion still need data.

Review

Search optimization starts with clear product definitions and page structure. Its impact still requires published-site evidence.

Ask about the work ↓
04

Proposal workflow · Generation and review

Standardize varied source inputs and proposal outputs; three iterations used for 50 proposals.

Result & boundaryI used the workflow to complete 50 proposals and made it available to business colleagues. Client content remains private.

Explore decisions and architectureClose decisions and architecture

The problem

Source material is scattered across text, images, documents, spreadsheets, video and messages, creating repeated preparation and layout work.

My responsibility

Define templates and input requirements, support varied sources, and arrange outline review before generation and final content/render checks afterward.

Implementation

Fix the template and required information first, then accommodate source formats. Feed failed inputs and corrections into three workflow iterations.

Trade-offs

Review facts and the technical outline before generation; check content and layout afterward so polished formatting cannot conceal factual errors.

Architecture

Source checks → Outline and detail review → Generation → Content and render checks

Evidence and boundaries

I used the workflow to complete 50 proposals and made it available to business colleagues. Client content remains private.

Review

Templates and checks are reusable. The facts in each new proposal still need verification.

Ask about the work ↓
05

ZJ Agent · Collaboration trial

Develop and deploy a cross-team workspace; the trial did not improve workflow efficiency.

Result & boundaryThe trial did not resolve workflow-efficiency problems.

Explore decisions and architectureClose decisions and architecture

The problem

Attempt to align cross-team project information, ownership and collaboration actions.

My responsibility

Develop, deploy and organize the trial from late June to late July 2026, covering project registers, responsibility records, audit events and Agent proposals.

Implementation

Gather trial feedback against actual work; identify a mismatch between multi-role staff and the departmental model.

Trade-offs

I organized collaboration around a department hierarchy before validating one concrete task. That was the main design error.

Architecture

Project ledger → Responsibility records → Audit events → Agent proposals

Evidence and boundaries

The trial did not resolve workflow-efficiency problems.

Review

Next time, test one collaboration task with clear participants, actions and responsibility before expanding the Agent.

Ask about the work ↓

02 / Method

Choose by evidence.

I choose the direction. Practice changes what I can do.

AI · Technology · Team
  1. Explore

    Find problems, connect disciplines and identify work worth pursuing.

  2. Build

    Make my own decisions and combine software, hardware and design to build real work.

  3. Verify

    Test the results through experiments, real use and feedback, then revise my judgement.

  4. Reuse

    Develop reusable tools, standards and methods, extending their reach through AI, technology and teams.

What works becomes the starting point for the next challenge.

Assessing knowledge gaps

First assess what the problem needs, then decide how deeply to learn. Use AI to broaden exploration; keep my own judgement and verify through practice. Focus effort on current obstacles, material risks and capabilities that matter, while leaving room for curiosity.

How this works in engineering

Define inputs, states, outputs and exceptions. Test each link, then record failures and fixes.

  1. Define the problem

  2. Check the inputs

  3. Test each link

  4. Review and revise

Explore technical connections
Read technology relationships (49)
  • AI applications LLM
  • AI applications Coding Agent
  • AI applications Harness
  • AI applications Prompt / system design
  • AI applications RAG / knowledge base
  • AI applications Python
  • AI applications Vue
  • AI applications React
  • Computer vision OpenCV
  • Computer vision YOLO
  • Computer vision ONNX
  • Computer vision PyTorch
  • Computer vision DINO
  • Computer vision SAM
  • Computer vision 3D / mocap
  • Computer vision Python
  • Computer vision C# / .NET
  • Full-stack engineering Vue
  • Full-stack engineering React
  • Full-stack engineering Next.js
  • Full-stack engineering TypeScript
  • Full-stack engineering Vite / ESM
  • Full-stack engineering Node.js
  • Full-stack engineering Design & IA
  • Platform delivery C# / .NET
  • Platform delivery Python
  • Platform delivery FastAPI
  • Platform delivery SQL / cache
  • Platform delivery Linux / deploy
  • Platform delivery Docker
  • Platform delivery Git / collaboration
  • Platform delivery Testing / acceptance
  • C# / .NET FastAPI
  • Python FastAPI
  • Vue TypeScript
  • React Next.js
  • React TypeScript
  • LLM Coding Agent
  • Coding Agent Harness
  • ONNX YOLO
  • Linux / deploy Docker
  • Harness Testing / acceptance
  • Design & IA Vue
  • 3D / mocap ONNX
  • DINO SAM
  • DINO ONNX
  • SAM YOLO
  • DINO PyTorch
  • SAM PyTorch
References & general-purpose tools

Small tools for checking data and interfaces, running in your browser.

Back to capabilities and practice ↑

03 / About

Experience & background.

Work

Algorithm engineer at Zhongjun Vision, Suzhou, June 2026 – present. I temporarily also lead the hardware group for project needs. My long-term focus is software development and algorithms.

Education

BSc in Intelligent Science and Technology, Xi’an University of Science and Technology, September 2022 – June 2026.

AI in practice

How I use AI in development

From GPT-3.5 to coding agents and harness engineering, I use AI to break down tasks, assist development and verify results. I own technical decisions, review and lessons learned.

Independent personal experience

Anx Day

— 28 days of practice

A self-directed period of building: turn real problems from work and learning into useful pieces of work, and keep a record of iterations, choices and reflection.

Explore Anx Day (Chinese) →

04 / Ask

Ask specifically.

Answers use published sources. Missing evidence stays missing.

Scope: all published sources

No conversation is saved in this browser. Questions and answers are excluded from site statistics.0/800

Anonymous site interaction statistics. Inputs are excluded.